Tracking Multiple Surface Vessels with an Autonomous Underwater Vehicle: Field Results
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
©2020 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or 1 redistribution to servers or lists, orreuse of any copyrighted component of this work in other works. https://doi.org/10.1109/JOE.2020.3015415 Wolek, A., McMahon, J., Dzikowicz, B. R., and Houston, B. H. (2021). Tracking Multiple Surface Vessels with an Autonomous Underwater Vehicle: Field Results. IEEE Journal of Oceanic Engineering. Tracking Multiple Surface Vessels with an Autonomous Underwater Vehicle: Field Results Artur Wolek∗ , James McMahon† , Benjamin R. Dzikowicz† , and Brian H. Houston† ∗ Department of Mechanical Engineering and Engineering Science, The University of North Carolina at Charlotte, Charlotte, NC Email: awolek@uncc.edu † Code 7130, Physical Acoustics Branch, Naval Research Laboratory, Washington, D.C. Abstract—This paper describes the development and testing length of the array (tens or hundreds of meters) reduces ma- of a passive sonar, multi-target tracker, and adaptive behavior neuverability and increases system complexity. Alternatively, that enable an autonomous underwater vehicle (AUV) to de- rigid hull-mounted sonars have a shorter-aperture but are more tect and actively track nearby surface vessels. A planar hull- mounted hydrophone array, originally designed for active sonar, compact. For example, AUVs have been outfitted with cross- is repurposed for passive sonar use and provides acoustic data track arrays (e.g., embedded in the wings of a glider [15]), to a time-delay-and-sum beamformer that generates multiple rigid planar arrays [9], line arrays [4], or volumetric arrays angle-only contacts. A particle filter tracker assimilates these that facilitate range estimation [2], [3], [10]. Here, a tracking contacts with a single-hypothesis data association strategy to system is developed that repurposes the receive array and data estimate the position and velocity of targets. Summary statistics of each track are periodically reported to an onboard database, acquisition system of a hull-mounted active sonar [21] for along with qualitative labels. To improve tracking performance, passive sonar use. Although the sonar was not purpose-built detections trigger an adaptive behavior that maneuvers the AUV for passive sensing, the proposed approach enables detect- to maintain multiple targets in the field of view by minimizing ing and tracking nearby vessels without requiring additional the worst-case aspect angle deviation from broadside (across hardware. The paper describes the acoustic signal processing all targets). The tracking system is demonstrated through at- sea experiments in which a Bluefin-21 AUV adaptively tracks chain, multi-target tracker, and a novel adaptive behavior that multiple surface vessels, including another autonomous platform, maneuvers the AUV to improve tracking performance. in the approaches to Boston Harbor. AUVs have historically been deployed along pre- Index Terms—Autonomous underwater vehicle (AUV), adap- programmed routes to record acoustic data for later analysis tive behavior, multi-target tracking, passive sonar. [2], [3], [22]–[26]. However, adaptive behaviors that direct AUV motion in real-time are increasingly common. Acous- tically detecting the presence of a surface vessel can trigger I. I NTRODUCTION surfacing and diving [11] or avoid collisions [10]. Behaviors Detecting and tracking moving targets with a passive sonar can maneuver an array to resolve a target’s port/starboard onboard an autonomous underwater vehicle (AUV) enables position ambiguity [8], [9], [27], maintain a target broadside ship traffic monitoring [1], undersea surveillance [2]–[5], and [9], [27], [28], or close range to a target [7], [8], [29]. marine life observation [6]–[8]. Target tracking also provides External moored passive acoustic systems can guide an AUV situational awareness to aid in mitigating collisions [9]–[12]. to intercept a target [30]. Teams of AUVs can cooperatively Passive tracking is challenging because of the complexity of localize a static source [18], track a moving target by encir- underwater sound transmission and the presence of high levels cling it [31], or collectively keep a moving target broadside of background noise and interfering sources [13]. AUVs with [27]. Behaviors have also been developed for systems that passive (or active) sonar often undergo extensive modifications detect multiple targets; however, only one target is tracked to reduce self-generated noise that arises from mechanical at a time. For example, by minimizing the trace of a target sources (e.g., actuators, fans) [14] or from the hydrodynamic estimate covariance matrix of the track with a highest track flow around the vehicle [15]. Tracking is further complicated score [32] using myopic heading control [33] or nonmyopic by the fact that the source strength and frequency spectrum of receding horizon control [32]. Switching from a loiter to a surface vessels is unknown [16], ship noise radiates asymmet- broadside behavior around the strongest contact has also been rically [16], and ship noise varies with ship speed and turn rate demonstrated [28]. Rather than tracking individual targets, [17]. To measure the relative bearing of a moving noise source, this work demonstrates balancing tracking of multiple targets an array of hydrophones attached to a long flexible cable is simultaneously by maneuvering the AUV to minimize the often towed behind an AUV [18]–[20]. Towed array sonars are worst-case aspect angle deviation from broadside (across all ideal for long range and high resolution measurements, but the targets).
2 Conventional pulsed [28], [32]–[36] or continuous active sonar data that allows efficient track exfiltration to command [37] sonar tracking experiments with AUVs are closely related and control (C2) entities via low bandwidth communications, to experiments with passive sonar. However, active multistatic such as acoustic modems. systems generate range measurements and require different The paper is organized as follows. Section II describes signal processing and infrastructure (e.g., low-noise sensors the target motion model, measurement model, and particle and data acquisition to mitigate electromagnetic or acoustic in- filter for single-target tracking. Section III overviews multi- terference). Furthermore, sensor networks with active elements target tracking, the track-labeling scheme, and the adaptive may not be suitable for the single-vehicle missions envisioned behavior. Section IV describes the vehicle platform, autonomy here. architecture, and the experiment design. Section V discusses Many passive tracking experiments with AUVs utilize cal- field results. Lastly, Sec. VI concludes the paper. ibrated acoustic sources [4], [18], [38], simulated bearing sensors [29], or acoustically tagged targets [8] as a proxy for actual ship noise. Prior experiments that processed ship II. S INGLE -TARGET T RACKING noise in real-time focused on detection [10], [11] and bearing This section describes the single-target tracking approach, measurement [4], [9] for a single target. In [39] a wave glider including: the target motion model for surface vessels, the detects, tracks, and classifies multiple surface vessels with a passive sonar sensor model that generates aspect angle mea- towed volumetric hydrophone array, but it does not actively surements, and the particle filter that estimates the position maneuver in response to targets. This paper adds to the limited and velocity of a target. First, some prerequisite notation is literature on passive tracking of actual platform noise in real- introduced. time by an AUV. Our approach detects multiple noise sources as peaks of an energy amplitude versus aspect angle curve that is generated by a time-delay-and-sum beamformer. Contacts A. Notation are assimilated by a multi-target tracker to estimate the state Bold lowercase letters denote real-valued vectors, e.g., x ∈ of surface vessels, and the AUV maneuvers in response. Rn , and bold uppercase letters denote real-valued matrices, Techniques for inferring target motion from acoustic mea- e.g., P ∈ Rm×n . Let x ∼ N (µ, σ 2 ) denote a normally surements are well known [40]. State estimation involves distributed random assimilating noisy measurements to estimate the position and √ variable with probability density 2 function g(x; µ, σ 2 ) = (σ 2π)−1 exp(−1/2 [(x − µ)/σ] ), where µ is velocity of one or more targets. Various filtering techniques the mean, and σ 2 is the variance. Similarly, let x ∼ N (µ, Σ) have been implemented onboard AUVs for single-target track- denote a normally distributed random vector with vector mean ing. These include alpha-beta filters for bearing stabilization µ and covariance matrix Σ. [4], particle filters [8], [31], or extended Kalman filters [27]. Kalman filtering can incorporate state-dependent noise statis- tics of hydrophone arrays [18]. Multi-target tracking (also B. Discrete-Time Constant Velocity Target Model single-target tracking in clutter) is challenging because it The discrete-time motion of a target surface vessel is rep- requires a mechanism for data association, i.e., pairing mea- resented by a constant velocity second-order kinematic model surements with tracks [13]. Most work involving real-time data subject to position and velocity disturbances. The target state association onboard AUVs has been conducted at the NATO T at the kth timestep is xk = [nk ṅk ek ėk ] , where nk and Undersea Research Center and considers data association for ek are the northing and easting of the target, respectively, and a single-target with a Kalman filter [33] or for multiple targets ṅk and ėk are the target velocities. The state equation is [41, with a particle filter [35] using an active multistatic network. p. 269] Our work is related; however, we consider data association for multi-target tracking using a passive acoustic system and 1 Tk 0 0 nk−1 (wn )k a particle filter. Our formulation accounts for passive sonar 0 1 0 0 ṅk−1 (wṅ )k sensor limitations, such as endfire saturation and port/starboard xk = 0 0 1 Tk ek−1 + (we )k , (1) ambiguity. We devise a rules-based approach to label tracks 0 0 0 1 ėk−1 (wė )k qualitatively so that they provide contextual information to | {z } | {z } | {z } Fk xk−1 wk other onboard processes. The contributions of this paper are the development and where Fk is the state-transition matrix, and Tk = tk − tk−1 experimental at-sea demonstration of: (1) an onboard passive is a fixed sampling time required to record a frame of data acoustic signal processing chain for multi-target detection and from the passive sonar. The zero-mean Gaussian random aspect angle measurement using an AUV with a rigid hull- disturbance vector wk ∼ N (0, Wk ) has covariance matrix mounted array; (2) a tracking framework that estimates multi- Wk = diag([σp2 σv2 σp2 σv2 ]T )Tk , where σp and σv are process ple targets while considering passive sonar sensor limitations noise parameters. The speed of the target and providing qualitative track labels; and (3) an autonomous q behavior that balances tracking of multiple targets by maneu- vk = ṅ2k + ė2k (2) vering to maintain them in the field of view. The onboard signal processing demonstrated in these experiments exhibits is bounded by a minimum and maximum speed, i.e., vk ∈ V, a data reduction of several orders of magnitude from the raw where V = [vmin , vmax ].
3 are detected if rk ∈ R, where R = [rmin , rmax ] defines the sonar’s sensing range. (A maximum range rmax is assumed for target estimation; however, in principle, the range of a passive sonar is unlimited, as it depends on the acoustic source level and the presence of noise.) D. Likelihood Function In Bayesian target estimation, the likelihood function is the relative probability of a measurement (i.e., the aspect angle y ∈ [0, π]) for all target states (i.e., positions and speeds x ∈ R4 ). However, the likelihood is not a probability density function—it does not necessarily integrate to unity across the target state space. The likelihood is designed from the measurement equation (6) and other practical considerations Fig. 1: Geometry of the passive sonar model with an axially oriented on the target speed and sensor range limits as follows. hull-mounted hydrophone array. The vehicle has blind spots at endfire The measurement space of aspect angles is partitioned into (i.e., in front and behind). The effective field of view angle is ϕ. The three regions Ef ∪ F ∪ Ea = [0, π], where Ef = [0, αf ] is sonar reports an aspect angle α that is port/starboard ambiguous. the fore endfire region, F = (αf , αa ) is the field of view, and Ea = [αa , π] is the aft endfire region. Measurements in endfire, Ea or Ef , are assumed to be saturated and handled as a special C. Passive Sonar Sensor Model case by the multi-target tracker (described later in Sec. III). The observing AUV (also referred to as the ownship) has States that lie outside the speed or range interval, V or R, state pk = [nok eok ψko ]T at time tk , given by its northing nok , respectively, have zero likelihood. To express this condition, easting eok , and heading ψko . The bearing of a target relative to an indicator function IA : Rd → {0, 1} maps a d-dimensional the ownship is vector x ∈ Rd to 1 if x ∈ A ⊂ Rd and to zero otherwise. For βk (xk , pk ) = atan(∆Ek /∆Nk ) , (3) a given target and ownship state, the speed, aspect angle, and range, are computed using (2), (4), and (7), respectively. The where ∆Ek (xk , pk ) = ek − eok and ∆Nk (xk , pk ) = nk − nok product of the indicator function and a Gaussian distribution are relative distances. The target bearing and ownship heading about the measurement is the likelihood are measured clockwise from north. The aspect angle of the target L(y|x; p) = IV×R (v, r)g(α; y, σy2 ) , (8) αk (xk , pk ) = d] (ψko , βk (xk , pk )) (4) where g(·) is given in Sec. II-A. The likelihood (8) is the is the conical angle measured relative to vehicle’s heading (see basis for the measurement update step of the particle filter, as Fig. 1), where d] (·, ·) : S × S → [0, π] returns the shortest un- described next. signed angular distance between two input angles. According to (4), if a target is fore (i.e., in front of the ownship) the aspect E. Single-Target Tracking using a Particle Filter reads zero, and if the target is aft (i.e., behind the ownship) it reads π. The measurement is port/starboard ambiguous and Target tracking estimates the state xk of a target with does not distinguish between a noise source on the port or dynamics (1) from a series of noisy measurement y1:k = starboard sides of the axial array. Define the possible port {y1 , . . . , yk } obtained via (6) at times {t1 , . . . , tk }. The and starboard bearings for a given aspect angle and ownship estimated target state is a random vector characterized by heading as the posterior probability density function (pdf) p(xk |y1:k ). A recursive Bayesian estimator infers the target state from βport (α, ψ) = ψ − α and βstar (α, ψ) = ψ + α . (5) measurements and the target motion model using Bayes Rule The measured aspect angle is corrupted by additive zero-mean [42, Ch. 15]: Gaussian noise qk ∼ N (0, σy2 ) and subject to saturation limits L(yk |xk )p(xk |y1:k−1 ) due to endfire. Thus, the aspect angle sensor reports p(xk |y1:k ) = , (9) p(yk |y1:k−1 ) yk = sat(αk (xk , pk ) + qk ; αf , αa ) , (6) where p(xk |y1:k−1 ) is the target pdf after the motion update where sat(x; xlb , xub ) is a saturation function that bounds (computed by applying the Chapman-Kolmogorov equation a scalar input x between a lower bound xlb and an upper with the state transition probability p(xk |xk−1 ) from (1) and bound xub . The constants αf = (π − ϕ)/2 and αa = π − αf the prior p(xk−1 |y1:k−1 )), L(yk |xk ) is the likelihood of the are the minimum (fore) and maximum (aft) aspect angles, measurement (8), and p(yk |y1:k−1 ) is a normalizing constant. respectively, where ϕ < π is the effective field of view angle The recursive Bayesian estimator applies to nonlinear systems, (see Fig. 1). Targets at a range but since closed form expressions for (9) only exist for some q special cases (e.g., with linear dynamics and measurement rk = ∆Ek2 + ∆Nk2 (7) equation), the particle filter is adopted here.
4 A particle filter represents the target pdf p(x|y) by a d) Roughening [43]: Compute the maximum differ- set of N particles x(i) ∈ R4 and weights w(i) ∈ R, for ence across each particle dimension mk (l) ← (i)∗ (j)∗ i ∈ {1, . . . , N }. The filter recursively applies (9) as described maxi,j∈{1,...,N } |xk (l)−xk (l)|+mmin , where in the following PARTICLE F ILTER algorithm, adapted from (i)∗ mk (l) and xk (l) denote the lth element of mk [42, Sec. 15.2]. At the beginning of the kth timestep, particles (i)∗ and xk , respectively, for l ∈ {1, 2, 3, 4}, and are distributed according to the prior pdf p(xk−1 |y1:k−1 ) with mmin is a small additive constant (for numerical uniform weights. The motion update propagates each particle stability). Roughen each particle i ∈ {1, . . . , N } with the motion model (1) to approximate p(xk |y1:k−1 ). (i) with noise ∆xk ∼ N (0, Kdiag[mk ]), where K is Then, the measurement update adjusts the weight of particles a roughening gain, to obtain the posterior particles according to the likelihood L(yk |xk ) (8). To avoid maintaining (i) (i)∗ (i) xk ← xk + ∆xk . low-likelihood particles, a re-sampling step redistributes the e) Compute statistics: Obtain the mean and first-order N particles with equal weight according to the posterior pdf moment of the posterior particle distribution: p(xk |y1:k ). However, if the region of the state space where N the pdfs p(xk |y1:k−1 ) and p(xk |y1:k ) have significant value 1 X (i) do not overlap, then the resampled particle set may contain x̂k = x (10) N i=1 k only a few unique particles. This sometimes leads to an implementation issue called sample impoverishment, where and N all particles collapse to a single value [42, Sec. 15.3]. To X (i) (i) Pk = xk (xk )T − x̂k x̂T k . address this issue, a roughening step perturbs the particles i=1 after re-sampling to maintain particle diversity [43]. F. Track Labeling Algorithm: PARTICLE F ILTER To provide onboard decision-making processes with an 1) Initialize N particles: At the initial timestep k = 0, use indicator of the status of existing tracks, a summary report the measured aspect angle y0 and vehicle heading ψ0o to is periodically posted to a central database on the AUV. The compute the port and starboard bearings, βport and βstar , report includes the time since each track was created, the target respectively, using (5). Then, for i ∈ {1, . . . , N/2}: state mean and covariance, and one of the following labels: a) Generate a random speed v, range r, and course • spawned: initialized from a new measurement χ, sampled uniformly from V, R, and [0, 2π], • detected: target is confirmed respectively. • resolved: port/starboard ambiguity is resolved b) Generate a random bearing β sampled from the • converged: position uncertainty is below a threshold normal distribution N (βport , σy2 ). • endfire: target is in endfire c) Compute the particle state • expired: track is dissolved o n0 + r cos(β) The method by which tracks progresses through the above (i) eo0 + r sin(β) labels is described in the context of the multi-target tracker in x0 ← , Sec. III-C. v cos χ v sin χ III. M ULTI -TARGET T RACKING where (no0 , eo0 ) is the known ownship position. Multi-target tracking requires solving the data association 2) Repeat the above procedure for particles i ∈ {N/2 + problem, that is, associating M measurements to a set of P 1, . . . , N } while replacing βport with βstar in Step 1b. existing (or new) tracks. The global nearest neighbor (GNN) 3) For k ≥ 1: method commits to a particular measurement-to-track pairing (i) a) Motion update: For each particle xk−1 , generate a at each frame (referred to as single-hypothesis tracking) and (i) random process noise vector wk ∼ N (0, Wk ), is perhaps the simplest and most widely used method for and propagate the particle through the motion data association [45, Sec. 6.4]. The joint probabilistic data (i)− model (1) to obtain xk . association filter (JPDAF) [45, Sec. 6.6.2] makes multiple (i)− association hypotheses between each measurement-to-track b) Measurement update: For each particle xk , com- (i) (i)− pair that are weighted to form a combined target estimate. pute the unnormalized weight w̃k ← L(yk |xk ) using (8). Then, normalize the weights as wk ← (i) In multiple-hypothesis tracking (MHT) [45, Sec. 6.7], a tree (i) PN (j) of hypotheses across a finite number of past frames forms the w̃k / j=1 w̃k . target estimate. All of these strategies decompose multi-target c) Resampling [44]: Accumulate the weights into a (i) Pi (i) tracking into a series of single-target tracking problems which cumulative sum vector ωk = j=1 wk , such (1) (1) (N ) may use, for example, Kalman or particle filters. Alternatively, that ωk = wk and ωk = 1. For each some filters eliminate data association altogether by combining i ∈ {1, . . . , N }, draw a uniformly random number particle filtering with finite-set statistics, as in the hybrid s ∈ [0, 1], and increment the integer q = 1 by Bernoulli/Poisson filter [46] or the probability hypothesis (q) ones until ωk > s. Set the resampled particle density (PHD) filter [47]. For a survey of particle filter based (i)∗ (q)− xk ← xk . methods in multiple-target tracking refer to [48].
5 The GNN method was adopted here because of its ease of spawned or detected, or cmp = −L(ym |(x̂last )p ; pk ) if a implementation and concerns that more sophisticated multi- track has status resolved, converged, or endfire. Both target tracking methods (e.g., MHT, JPDAF, PHD) may have costs have the same minimum and maximum value, but the prohibitive computational requirements — in this work, the difference is that the former is applied during initial stages of tracker software resides on an autonomy computer that also the track when the estimated target state is unreliable (because shares computational resources with the sonar processing and the port/starboard ambiguity has not yet been resolved). The vehicle control software. A limitation of the GNN method is assignment problem is then that it may perform poorly in high clutter environments, and it 0 P X M 0 is most suitable for tracking a few well-spaced targets in low X min cmp xmp , (11) noise scenarios. p=1 m=1 In our approach, data association proceeds in stages: a P 0 gating process identifies likely measurement-to-track pairs, X subject to xmp = 1 for all m ∈ {1, . . . , M 0 } , a pre-processing stage forms new tracks and makes as- p=1 signments where no association ambiguities exist, and the M 0 assignment stage pairs the remaining unassigned measure- X xmp = 1 for all p ∈ {1, . . . , P 0 } , ments and tracks. Lastly, track maintenance labels tracks and m=1 dissolves them, if necessary. This process is executed by and xmp ∈ {0, 1} , the M ULTI - TARGET T RACKER algorithm summarized in the upper left block within Fig. 2. (A full explanation of the where the decision variable xmp = 1 represents the assignment remaining processes in Fig. 2 will follow shortly.) of measurement m to track p. Applying the Hungarian algo- rithm [49] to solve (11) yields the minimum-cost assignment. For each assigned measurement-track pair, Step 2 of the A. Gating and Pre-processing PARTICLE F ILTER algorithm is executed. If the assignment A gate is a region in the measurement space that can problem is solved with more measurements than existing reasonably be associated with a given track. For each track p ∈ tracks, M 0 > P 0 , then M 0 − P 0 measurements are left {1, . . . , P }, a gate Gp = [(ylast )p −δ/2, (ylast )p +δ/2] ⊂ [0, π] unassigned. However, because of the pre-processing step, these is defined, where (ylast )p is the last aspect measurement to be measurements are within a track’s gate that was updated by associated with track p (with aspect corrected for any heading the assignment problem. Rather than spawning a new track changes), and δ is the angular width of the gate. The binary in close proximity to an existing one, these measurements are M ×P gate matrix G identifies measurements m ∈ {1, . . . M } ignored. Conversely, if there are more unassigned tracks than that are within each gate by assigning a value of one to the measurements, P 0 > M 0 , then only a subset of the tracks are pth row and mth column if ym ∈ Gp and a value of zero updated. otherwise. Saturated measurements in endfire are momentarily ignored. C. Track Maintenance If targets are well separated, and there is little clutter, some As discussed in Sec. II-F, the tracker periodically labels measurement-track pairs may be associated at this stage. If tracks and reports summary statistics to an onboard database. the mth row of G contains all zero entries, the mth measure- A rules-based scheme updates the track status according to ment spawns a track using Step 1 of the PARTICLE F ILTER the T RACK S TATUS U PDATE algorithm below (summarized in algorithm, and measurement m is not considered further for Fig. 3). Tracks are initially labeled spawned, and estimation data association. If the pth column of G contains all zero begins even before a target is detected. The conventional GNN entries, then no measurement is within track p’s gate, and approach of using a heuristic “M/N rule” to confirm tracks the track is not updated at the current timestep. If the pth and a “ND consecutive misses” to dissolve tracks [45, Sec. column of G contains exactly one non-zero entry, then the 6.4] is adopted. After several measurements are associated corresponding measurement and track are associated, and Step to the track, its status is elevated to detected to indicate 2 of the PARTICLE F ILTER algorithm is executed. Thus, after confirmed target presence. Otherwise, if no measurements gating and pre-processing, there remain M 0 ≤ M unassigned are associated with a track after several frames, it becomes measurements and P 0 ≤ P unassigned tracks. Each unas- expired. The particles in the filter initially have a bimodal signed track has two or more unassigned measurements within distribution due to the port/starboard ambiguity. However, if a its gate. sufficient number of particles transition from port to starboard, or vice-versa, the track is labeled resolved. Furthermore, B. Assignment Problem if the track uncertainty reduces below a threshold, its status An assignment problem is formulated to pair the M 0 mea- becomes converged. Tracks that predict a target in the fore surements with P 0 tracks leftover after the pre-processing step. or aft blind spot of the vehicle are labeled endfire, and Let m ∈ {1, . . . , M 0 } index the unassigned measurements T RACK S TATUS U PDATE predicts when they return into view. and p ∈ {1, . . . , P 0 } index the unassigned tracks. The last Algorithm: T RACK S TATUS U PDATE state estimate associated with the pth track is (x̂last )p . The At the kth timestep, after executing the initial filtering steps cost of assigning measurement m to track p is computed of the M ULTI -TARGET T RACKER algorithm (described in either as cmp = −g(ym ; (ylast )p , σy2 ) if a track has status Sec. III-A and III-B), do the following for each track:
6 Autonomy Computer pBearingTracker Nav. Data, Tracker Multi-Target MultiVehicleTracker Tracker Tracker Reports pHelmIvP Status Track 1 Reports ... Commands BHV_ConstantAltitude BHV_ConstantDepth Particle Filter Track P Track Track Aspects Data BHV_Waypoint Nav. Data Assignment Pre-Process Gate Buffer MOOS Database BHV_MVKeepBroadside BHV_MTKeepBroadside Contacts Unassigned Modified Gate Synchronized Measurements Gate Matrix Matrix Nav./Contacts Contacts Commands iFrontSeat Nav. Data HDD Standard Payload Reporting Frequency Binning Interface Vehicle Control Computer Peak Detection Beamforming Data Acquisition Smoothing Impulse Filter Raw Acoustic Computer HDD Passive Onboard Data Fig. 2: Process flow diagram illustrating computers and software components of the onboard tracking system. The Passive Onboard signal processing chain accepts raw acoustic data (pressure time series for each hydrophone in the array) and generates up to Mmax contacts. Each contact consists of an aspect angle, total energy, and energy binned by frequency. Contacts are posted to the onboard database and written to a log on the hard disk. The M ULTI -TARGET T RACKER algorithm is encapsulated in a MOOS application, pBearingTracker, that provides the contacts synchronized with interpolated navigation data held in a buffer. The adaptive behavior, BHV MTKeepBroadside, resides in the pHelmIVP application and sends heading, depth, and speed commands to the vehicle. spawned updates c) resolved: Compute the semi-major axis of the expired updates 2σ error ellipse dellipse [50]. If dellipse < dconverge , detected set the status to converged. measurement assigned d) endfire: If a measurement is associated after a resolved track has passed through endfire, set the status to endfire detected. converged Fig. 3: State machine diagram for updating the track status. 3) If the status is resolved, converged, or endfire: a) Propagate the state and the covariance to the cur- 1) Determine the number of measurements MF assigned rent time, using the linear update: x̂k = Fk x̂k−1 to the track over the last F frames. If MF = 0, then set and Pk = FT k−1 Pk−1 Fk−1 + Wk−1 . The esti- the status to expired. mated target state x̂k is used by the assignment 2) If no measurement was assigned to the track (at the problem, as discussed in Sec. III-B. current timestep) proceed to Step 3. Otherwise, execute b) Predict the aspect angle αcur = α(x̂k , pk ) using one of the following (depending on the current track (4) and determine if the target is fore or aft. status): Compute the probability that the target is in end- R R a) spawned: If MF ≥ M F , set the status to fire Pendfire = R B p(r, β)rdrdβ using Gauss- detected, where M F is a user-defined integer. Legendre quadrature, where p(r, β) is a Gaussian b) detected: Compute the estimated target bear- pdf of the target’s position in polar coordinates, and ing β̂ ← β(x̂k , pk ) using (10) and (3). De- B ∈ {Bfore , Baft } is the predicted endfire region, fine the interval of port and starboard bearings, either Bfore = {βstar (α, ψk ) ∪ βport (α, ψk ) | α ∈ Bport = {βport (α, ψk ) | α ∈ [0, π]} and Bstar = Ef } or Baft = {βstar (α, ψk ) ∪ βport (α, ψk ) | α ∈ {βstar (α, ψk ) | α ∈ [0, π]}, respectively, with (5). Ea }. Let B ∗ = Bport if β̂ ∈ Bport , and B ∗ = Bstar c) If Pendfire ≥ P , then set the status to endfire. otherwise. Determine the bearing of each particle d) If the status is endfire, and there exists a β (i) using (3), and let Nagree be the number of saturated measurement in the same endfire region particles with β (i) ∈ B ∗ , for i = {1, . . . , N }. If (Baft or Bfore ), then mark the track as updated with Nagree > Nresolve , set the status to resolved, an empty measurement (to prevent the track from where Nresolve is a user-defined integer. dissolving as it passes through endfire).
7 D. Multi-Target Keep Broadside Behavior experiment to test the tracking system and adaptive behaviors. The multi-target tracker generates a set of tracks repre- senting estimates of surface vessels or other noise sources. Suppose that a subset of these tracks, indexed by q ∈ A. Platform and Autonomy Architecture {1, . . . , Qk }, have a status of either detected, resolved, The experiment was conducted with a propeller-driven, 21- or converged at the kth timestep. Let αq (ψ) be the aspect inch diameter AUV developed by the Naval Research Labora- angles predicted for the target in the qth track (i.e., computed tory and Bluefin Robotics (see Fig. 5a). The tracking system using (4) and (10)). The multi-target keep broadside behavior utilizes three computers: (1) a vehicle control computer that commands the heading ψk∗ that minimizes the worst-case interfaces with hardware for guidance, navigation, and control aspect angle deviation from a broadside aspect (α = π/2). (GNC), using Huxley [51]; (2) a data acquisition (DAQ) Keeping target aspects close to broadside improves measure- computer that manages the passive sonar and records acoustic ment quality and prolongs the duration that targets remain in data; and (3) an autonomy computer for mission planning, the field of view. To compute ψk∗ , discretize the heading into vehicle behavior control, acoustic data processing, and multi- Nψ admissible values to form a set of candidate headings target tracking. The interactions between these computers are Ψ = {0, 1/(Nψ − 1)2π, . . . , Nψ /(Nψ − 1)2π}. Then, for each summarized in Fig. 2. The Mission Oriented Operating Suite candidate heading, predict the aspect angle αq (ψ) of each with Interval Programming (MOOS-IvP) robotic middleware target and evaluate the worst-case aspect-angle deviation from [52] provides a central database, the MOOSDB, where appli- broadside to determine the minimizing value cations exchange messages (time-stamped data consisting of ψk∗ = argmin max |π/2 − αq (ψ)| . (12) strings, doubles, or binary data). The autonomy architecture ψ∈Ψ q∈{1,...,Qk } | {z } adopts the backseat/frontseat control philosophy inherent in worst-case aspect angle deviation MOOS-IvP; Huxley performs critical GNC tasks (similar to the There exist pathological cases where a set of Qk arbitrarily po- driver of a car), and the autonomy computer provides high- sitioned targets cannot be simultaneously in view according to level commands (similar to a backseat passenger directing (12); however, since (12) considers only targets that are already a driver). These commands are heading, depth, and speed in view at the current time, these cases will not arise. In other setpoints issued periodically through the iFrontSeat application words, ψk∗ will not place targets into endfire that are not al- [53] using a Standard Payload Interface (SPI) protocol. The ready in endfire. Figure 4 illustrates a nominal case with three setpoints arise from multiple behaviors that perform special- targets where the outermost targets, one and three, determine ized tasks (e.g., maintaining altitude over a varying seabed, the worst-case aspect deviation from broadside. The optimal waypoint guidance). The pHelmIvP application resolves any heading ψk∗ places these outermost targets symmetrically about conflicts between behavior setpoints by considering each be- the broadside angle, i.e., |π/2 − α1 (ψk∗ )| = |π/2 − α3 (ψk∗ )|. havior’s priority and arriving at a compromise through interval If the tracking behavior drives the AUV outside the limits programming optimization. The vehicle is equipped with an of a predefined operational area, then a safety return behavior inertial navigation system that has a position error of less than commands the vehicle to transit back to within a threshold of 0.1% distance traveled. The custom autonomy framework has the origin before restarting the tracking behavior (see Fig. 4). a layered structure that includes reactive behaviors [9], task and motion planning [54], and goal reasoning [55]. Safety 25 Restart Return T1 Tracking Behavior 500 20 Northing (m) Behavior Depth (m) 0 15 Return -500 10 Circle T2 5 -500 0 500 Easting (m) (a) NRL’s Bluefin-21 AUV (b) Experiment geometry Operational Area Fig. 5: The Bluefin-21 AUV executed an adaptive tracking behavior within the test site shown as a dashed-line in panel (b), while two controlled targets traversed east-west and north-south tracks, labeled Fig. 4: Sketch of tracking and safety behaviors. The multi-target keep T1 and T2, respectively. broadside behavior commands the heading ψ ∗ that minimizes the worst-case deviation in aspect angle from broadside across all targets. A safety behavior returns the vehicle to the middle of the operational area if a predefined boundary is exceeded. B. Passive Sonar System A passive sonar signal processing application, called Pas- IV. E XPERIMENTAL A PPROACH sive Onboard, reads acoustic data from the data acquisition This section describes the AUV platform, software archi- computer and provides contacts to the MOOSDB. The pla- tecture, passive sonar system, and the design of the at-sea nar hydrophone array is a variation of the one originally
8 developed for active synthetic aperture sonar [21] and lies C. Experiment Design vertically in the center of a flooded section of the AUV. The vehicle was modified to operate with reduced self-noise. The experiment was conducted in Boston Harbor between Hydrophone signals are acquired and immediately written to August 13-15, 2018. The location of the test site was a 1 a hard drive in real-time by the DAQ computer. The Passive km box centered at 42◦ 23.6800 N, 070◦ 55.3950 W, just Onboard block in Fig. 2 provides an overview of the signal south of Nahant, MA in Broad Sound. The test site is adjacent processing chain. Before beamforming and frequency binning, to a shipping lane to the east, and numerous vessels were the raw data is subject to a non-linear impulse filter that observed operating within visual range during the experiment. segments the signal and clips peaks within each segment that Two surface vessels, the R/V Resolution (a 54-foot twin- have an amplitude greater than four standard deviations of engine catamaran) and the R/V Jamie Hannah (a 55-foot the noise within the segment. This filtering process reduces single-engine fiberglass research vessel), served as controlled interference from high amplitude, transient sounds, typically targets that traversed north-south and east-west tracks just 130◦ exhibit degeneracy. At these angles the mean processing chain, tracking system, and adaptive behavior. A angle is reported. The Fourier transform of the beamformed total of about five hours of acoustic and tracker data were signals is sent to a frequency binning algorithm that sums the collected over three days. Each day allowed for twelve hours amplitudes for sixteen frequency bins. Bins are logarithmically of ship time (dock-to-dock), and runs were scheduled in- spaced to compensate for an expected higher energy at lower between other AUV operations. Since this was the first in- frequencies. The energy in each frequency bin is recorded. water test of the system, filter parameter tuning was required. Tuning was conducted by reviewing data logs at the end of Following the binning, the total beam amplitudes are each day. However, due to time-constraints, a more careful smoothed across aspect angle, and the minimum and maximum analysis was only possible after the experiment concluded. total amplitude is determined for subsequent peak detection. Raw acoustic and tracker input data were recorded in a manner Smoothing applies a running average of nearest neighbors. that allows re-processing through the onboard software with Peak detection is performed on the sum across all bins modified parameter values. The vehicle’s adaptive behavior for each beam (total beam amplitude). The peak detection presented here responded to the real-time tracker output. algorithm searches for beam angles with an excess of energy Table I outlines the duration of each run and the frequency by examining the total amplitudes of each beam, as computed of the last track status before becoming expired. Tracks by the frequency binning algorithm. First, it determines if there with status of spawned are numerous and correspond to are any extreme values in a time series by rejecting time series spurious contacts where target presence was not confirmed where the difference between the maximum and minimum am- (these tracks quickly become expired). Tracks that attained plitudes exceeds the minimum amplitude. Then, a peak beam a status of detected or resolved correspond to intermit- is identified as a contact within a qualifying time series if its tent targets, whereas tracks with status converged represent total amplitude is greater than: (1) 0.1 (Amax − Amin ) + Amin , persistent targets that were tracked for longer durations (often where Amax and Amin are the maximum and minimum total between five to ten minutes). amplitudes across beams, respectively; (2) that of neighboring beams; and (3) other peaks within 3◦ of it. Contacts are sorted TABLE I: Summary of runs. The tally indicates the frequency of by total amplitude, with higher total amplitude peaks reported each track status for a given run. Tracks with converged status first. If more than Mmax contacts are identified, then only the represent persistent targets that were tracked for longer durations. first Mmax are reported. Run 1 2 3 4 5 Duration (min.) 31.2 48.8 83.9 56.5 79.8 A reporting algorithm writes the data acquisition clock time, spawned 85 107 107 135 252 contact angles, and their frequency bin data to a binary file for detected 20 53 58 52 80 resolved 10 27 27 24 33 future analysis. During real-time operation, the results are also converged 7 9 21 12 22 published to the MOOSDB for use by the tracker and other processes.
9 A. Passive Bearing-Only Beamforming Results before converging to the aspect measurements. This artifact is During the experiment, the maximum number of reported seen throughout Figs. 7b and 7d as the near-vertical portion contacts was set to Mmax = 3. Due to bandwidth limitations at the beginning of each track. The ability of the vehicle to of the hard drive bus, only every third signal (frame) was maintain targets in view is evident by the tendency of tracks processed. This section discusses an illustrative 11.8 minute to either converge to the broadside angle (for a single target— snapshot of the passive acoustic data, shown in Fig. 6. The data see Fig. 7d around three minutes) or remain close to broadside were re-processed to include Mmax = 4 contacts and the full (for multiple targets—see Fig. 7b around 35 minutes). data set including all frames. Figures 6a and 6b show the angle TABLE II: Multi-target tracker parameters. and amplitude, respectively, of the top four contacts. Saturation of the aspect angle near endfire is apparent in Fig. 6a. Due Parameter Symbol Value Units to angular degeneracy, as discussed in Sec. IV-B, the angular Number of particles per track N 10,000 — resolution is reduced away from broadside. Particles to clear ambiguity Nresolve 8,500 — Three instants (A, B, and C) are labeled for discussion in Effective field of view ϕ 115 deg Fig. 6a and Fig. 6b and are further examined in Figs. 6c– Position process noise σp 5 m 6e. Near time A, there are two lines of correlated contacts Speed process noise σv 0.1 (m/s) Measurement noise σy 1.5 deg in Fig. 6a that suggest the presence of distinct targets. At Target range interval R [0,1500] m around 800 seconds the contact positions swap. This indicates Target speed interval V [0,8] m/s a change in the relative amplitude of the two targets that may Particle roughening gain K 0.1 — occur because of changes in relative range, speed, direction, Minimum particle roughening mmin 0.1 — or a variety of other factors effecting the received signal Gate width δ 15 deg Measurements to declare detected MF 5 — amplitude. Indeed, the contacts near time A in Fig. 6b exhibit Number of past frames considered F 10 — similar aspect amplitudes. The contact angles reported at time Ellipse axis to converge dellipse 300 m A are distinct peaks of the relative beam amplitude curve (see Endfire probability threshold P 0.5 — Fig. 6c). However, in other cases, these peaks are more subtle, as in Fig. 6d, especially when there is a large disparity in the contact amplitudes, as in Fig. 6e. The inset plots in Fig. 6e of C. Multi-Target Tracker Results energy with frequency bin show a unique signature for each of the three contacts C.1, C.2, and C.3. While not considered A single track (no. 202 in Fig. 7d) is examined in Figs. 8 and here, the contact amplitudes and frequency signatures can 9 to illustrate the target state estimator. The particles in Fig. 8 aid data association. However, high variability in the ship’s give insight into the state of the particle filter as the track status signature spectrum is expected due to modal interferences in changes. Particles are uniformly dispersed in speed and course the propagation channel. when the track is spawned in Fig. 8a. As more measurements are assimilated, the target is confirmed and the status becomes detected. The track status is resolved once the particles B. Multi-Target Keep Broadside Behavior Results resolve the port/starboard ambiguity and estimate a north-east To illustrate the performance of the multi-target keep broad- course (Fig. 8b). Later, the filter correctly estimates that the side behavior, Runs 2 and 4 are examined in Fig. 7. The track- target is moving north and the uncertainty ellipse is sufficiently ing results presented here are re-processed with the parameters small so that the track status is converged (Fig. 8c). At the in Table II; however, the AUV responded to the tracker (with point of closest approach, the error in the estimated target state an earlier parameter set) in real-time during the experiment. is at a minimum value (near five minutes in Fig. 9). The safety Gray segments in the AUV path (Figs. 7a and 7c) that connect behavior redirects the AUV toward the middle of the box, and the perimeter of the test site to a safety return circle centered the track becomes expired. on the origin (10 m radius for Run 2, and 200 m radius for The track status update procedure (Sec. III-C) generally Run 4) correspond to the safety behaviors. The vehicle can worked well and indicated the status as expected when also be commanded to begin its mission by reaching the safety tracks were continuously updated with measurements. How- return circle. Colored line segments (Figs. 7a and 7c) indicate ever, the mechanism for recovering a track from endfire where the adaptive broadside behavior is active. Although the to detected status had mixed results. For example, in vehicle is programmed to loiter when no contacts are detected, Run 4 (Fig. 7d) track no. 28 is not recovered around eight it never enters this state because of the constant presence of minutes. Instead, a new track no. 40 is spawned. In different noise sources. The AUV tends to move along curved paths as circumstances, track no. 168 briefly enters endfire and is it tracks targets. recovered around 40 minutes (Fig. 7d). Recovering from Figures 7b and 7d show the aspect angle during the runs endfire is challenging because of errors introduced by the with an overlay of selected tracks that assimilated thirty or linear state update, degenerate measurements, and other newly more contacts. Since the particle filter initially spawns particles spawned tracks that complicate data association. Since the on both the port and starboard sides of the array, the mean adaptive behavior keeps targets broadside, most instances of estimated state is along the axis of the vehicle (i.e., fore tracks moving into endfire occurred when the safety behavior or aft) before resolving the ambiguity. Thus, the estimated was engaged (causing an abrupt heading change towards the aspect angle for each track begins either at 0 or 180 degrees origin).
10 Aspect Angle (deg.) 180 Max Contact 1 Contact 1 C.1 Aspect Amplitude Contact 2 C.3 Contact 2 150 Contact 3 Contact 3 Contact 4 B.1 Contact 4 120 A.2 B.2 C.1 90 C.2 B.1 60 A.1 A.1 A.2 C.2 30 A B C A B C.3 C B.2 0 0 600 700 800 900 1000 1100 1200 1300 600 700 800 900 1000 1100 1200 1300 Mission Time (s) Mission Time (s) (a) Angle of first four contacts. (b) Amplitude of first four contacts. Relative Beam Amplitude Relative Beam Amplitude Relative Beam Amplitude Mission Time A Mission Time B Mission Time C Energy 0.015 Energy 0.015 B.1 0.015 C.1 Energy A.1 Energy A.2 Energy B.2 Freq. Bin 0.01 Freq. Bin 0.01 Freq. Bin 0.01 Energy C.2 Energy Freq. Bin C.3 Freq. Bin 0.005 Freq. Bin 0.005 0.005 Freq. Bin 0 0 0 0 50 100 150 0 50 100 150 0 50 100 150 Aspect angle (deg.) Aspect angle (deg.) Aspect angle (deg.) (c) Relative beam amplitude and frequency con- (d) Relative beam amplitude and frequency con- (e) Relative beam amplitude and frequency con- tent of contacts at time A. tent of contacts at time B. tent of contacts at time C. Fig. 6: Snapshot of passive bearing-only beamformer results over 11.8 minutes from Run 3. Contacts are color-coded with red denoting the highest total amplitude contact, through blue and green, to the fourth lowest total amplitude contact in black. Three specific instants (A, B, and C) are highlighted with relative beam amplitude and frequency content shown. 600 180 54 45 95 400 40 150 25 79 Start 35 56 133 Mission Time (min.) Aspect (deg.) 200 120 Northing (m) 73 30 14 26 188 0 25 90 151 20 39 63 122 -200 End 60 21 15 16 10 -400 30 158 5 94 -600 0 0 -600 -400 -200 0 200 400 600 0 5 10 15 20 25 30 35 40 45 Easting (m) Time (min) (a) Path of the AUV (Run 2). (b) Aspect angle over time with overlay of selected tracks (Run 2). 600 180 55 56 1 End 50 400 150 79 45 135 213 1 Mission Time (min.) 40 Aspect (deg.) 200 120 Northing (m) 35 3 51 61 196 64 145 30 90 0 Start 25 202 41 198 20 60 57 137 -200 191 15 142 -400 10 30 96 168 5 6 -600 0 0 -600 -400 -200 0 200 400 600 0 5 10 15 20 25 30 35 40 45 50 55 Easting (m) Time (min) (c) Path of the AUV (Run 4). (d) Aspect angle over time with overlay of selected tracks (Run 4). Fig. 7: Experimental results illustrating the multi-target keep broadside behavior during Run 2 and Run 4. The behavior maneuvers the vehicle to maintain multiple targets in the field of view. Left: Path of the AUV with gray segments indicating safety behavior returning the AUV to the middle of the box after exceeding operating area limits. Right: Contacts are shown as gray circles in background, dashed lines indicate approximate saturation limits and the broadside angle. Superimposed colored lines are selected tracks that associated thirty or more contacts (with track number indicated). Shaded areas are time intervals where the AUV safety behavior was active. Data association and behavior performance is illustrated by examining two multi-target tracking scenarios in Fig. 10.
11 1000 1000 1000 500 500 500 Northing (m) Northing (m) Northing (m) 0 0 0 -500 -500 -500 -1000 -1000 -1000 -1000 -500 0 500 1000 -1000 -500 0 500 1000 -1000 -500 0 500 1000 Easting (m) Easting (m) Easting (m) 0 0 0 30 30 30 30 -6 0 -6 0 -6 0 -90 90 -90 90 -90 90 -12 20 -12 20 -12 20 s s s 150 150 50 50 50 50 180 180 180 (a) Track status spawned at time 1.10 sec. (b) Track status resolved at time 1.29 min. (c) Track status converged at time 5.75 min Fig. 8: State of the particle filter at three instants during which the AUV tracks the R/V Resolution (RES) over nearly seven minutes (corresponding to track no. 202 in Fig. 7d). Particles (orange dots) have planar position and velocity. Top: Particle positions are plotted along with the mean (black “+” marker) and 95% confidence region (black ellipse). AUV motion is indicated by the blue line, starting at the circular marker and ending at the square marker. Similarly, RES motion is indicated by the black line and markers. The gray fan-shaped area indicates the field of view of the AUV and magenta rays are aspect angle measurements at each instant. Bottom: Polar plot of speed and course of particles—maximum speed bound of 8 m/s is indicated by a black circle. Mean speed of particles is shown as a black “+” marker. The three instants Fig. 8a-8c have track status of spawned (initial measurement, possible target), resolved (target presence confirmed and port/starboard ambiguity resolved), and converged (position error ellipse axis below threshold). In addition to the controlled targets, other ships with AIS targets, other sources of interference, or the loudness of one transponders were operating in the vicinity (see Table III). target obscuring the peaks of other targets in the relative beam The estimated positions of tracked targets is indicated by the amplitude plot used for identifying contacts (recall that the confidence ellipses in Figs. 10a and 10c. The aspect angle number of contacts reported was limited to Mmax = 3). measurements are compared to groundtruth in Figs. 10b and Now, consider the second scene in Figs. 10c and 10d; there 10d. Figure 10a shows track no. 3 estimating the position are two controlled targets (RES and JAM), the tanker NOR, of RES as it makes a 90 degree course change. The AUV and the ferry NAT. Initially, the AUV tracks JAM (only) adapts its heading to compensate for the fluctuation in aspect with track no. 39. At about 12 minutes, the AUV begins angle from broadside (see Fig. 10b at 2.5 minutes). Near five to also track NAT. When the aspect angle measurements are minutes, track no. 28 develops for a unmanned surface vessel, not well separated, data association is more challenging and STE, that appears to be performing a survey near the test site. track no. 39 dissolves (spawning a new track no. 63 that At six minutes, the STE track is resolved, and the AUV continues to track JAM). Tracks no. 56 and 73 track NAT and maneuvers to maintain both tracks equidistant in aspect angle NOR, respectively. Since both NAT and NOR are beyond the deviation from broadside (refer to Fig. 10b). Shortly thereafter, maximum range assumed by the filter, their position estimates the passive sonar stops reporting contacts from RES, and track are constrained to be closer to the AUV (see Fig. 10c). no. 3 dissolves; the AUV tracks STE only. At this time, the NOR and RES follow different spatial trajectories but have a RES is over 1 km away from the AUV, whereas STE is much similar aspect angle history throughout the scene. However, closer. Throughout this eight minute portion of Run 2, only few contacts are reported for RES and it is not tracked. a few contacts are reported from JAM and many are initially Between about 17.5-19 minutes the AUV tracks three targets missing from STE. The intermittent nature of contacts may (NAT, NOR, and JAM). When NAT moves into endfire it be due to the changing relative geometry of the AUV and transitions to tracking JAM and NOR only.
12 225 Actual RES End Bearing (deg.) 180 Estimated 1000 135 90 t = 6 min. 45 AUV End 0 500 t = 5 min. 6 5 Northing (m) Speed (m/s) t = 4 min. 4 AUV Start 3 0 2 t = 3 min. 1 180 t = 2 min. -500 Course (deg.) 90 t = 1 min. 0 -90 -1000 RES Start -180 0 1 2 3 4 5 6 -1000 -500 0 500 1000 Mission Time (min) Easting (m) (a) Estimated bearing, speed, and course of the target compared to ground (b) Estimated position of target compared to ground truth. Refer to Fig. 8 truth. The estimated bearing is computed as the angle between the mean for explanation of symbols and line colors. The six orange ellipses are planar position of the particles and the AUV. After about 20 seconds the estimates of the target position (actual position indicated by the six black port/starboard aspect angle ambiguity is resolved and the correct bearing is circular markers). estimated. Fig. 9: Tracking performance over a nearly seven minute interval during which the AUV tracks the R/V Resolution (RES) (data correspond to track no. 202 in Fig. 7d and the particle filter snapshots in Fig. 8). GPS logs from the RES provide position over time that determine the actual speed and course. GPS positions are compared to AUV navigation logs to obtain the actual bearing. TABLE III: The acronym, name, maritime mobile service identity (MMSI) number, and type of surface vessels encountered in the small number of well separated targets. The onboard signal scenarios of Fig. 10. processing demonstrated in these experiments exhibits a data reduction of several orders of magnitude from the raw sonar Vessel Acronym: Name MMSI Type data that allows efficient track exfiltration to command and JAM: R/V JAMIE HANNAH 367534250 research vessel control (C2) entities via low bandwidth communications, such NAT: NATHANIEL BOWDITCH 367534240 ferry as acoustic modems. NOR: NOR EASTER 538002783 tanker However, some challenges were also encountered, including RES: R/V RESOLUTION 338112818 research vessel difficulty with data association in cluttered scenarios and STE: STEADFAST 338233113 Sea Machines USV spurious measurements creating a large number of short- lived tracks. The number of contacts reported by the passive beamformer was limited to three, and in several instances a VI. C ONCLUSION controlled target used in the experiment was not detected, The ability of a Bluefin-21 AUV to respond to platform even at close range. Ongoing work aims to develop additional noise and track multiple nearby surface vessels simultaneously behaviors for adaptive data collection, incorporate frequency using a hull-mounted active sonar repurposed for passive sonar and energy information for data association, and implement a use was demonstrated. The key components of the tracking more robust but computationally efficient multi-target tracking system include: (a) a passive sonar utilizing a peak detector algorithm. to identify local maxima of an energy amplitude versus aspect angle curve that is generated by a time-delay-and-sum ACKNOWLEDGEMENTS beamformer; (b) a tracker that assimilates the resulting angle- only contacts using a particle filter with single-hypothesis The work of A. Wolek was supported by the American data association; (c) a rules-based scheme that qualitatively Society for Engineering Education (ASEE) through the ASEE labels tracks to indicate their status (e.g., spawned, detected, postdoctoral fellowship program at the Naval Research Lab- port/starboard ambiguity resolved, endfire, converged); and (d) oratory. The work of J. McMahon, B. R. Dzikowicz, and B. an adaptive behavior that maintains multiple targets in the H. Houston was supported by the Office of Naval Research field of view by optimizing the vehicle heading. The system through the Naval Research Laboratory base program. We was tested near a busy shipping lane in Boston Harbor in thank Alain Berdoz and Harry Simpson for their role in August 2018. Reasonable performance was observed when organizing experiment logistics, and Mark Wilson and Patrick tracking a single target with low clutter or when tracking a Amy for their help in executing the field trials. The authors are
You can also read