Underwater Search and Tracking: A Technical Guide to Target Detection, Search Planning, and Multi-Target Tracking

1. From Undersea Pursuit to Underwater Search and Tracking

Popular fiction often portrays undersea pursuit as a fast chase. In engineering terms, the problem is much more demanding: how to reliably detect and continuously track a moving target in an ocean environment with weak light, limited communication, time-varying conditions, and heavy interference.

Three practical drivers make this problem important:

• Underwater infrastructure inspection. Subsea pipelines and cables require periodic inspection. Systems must detect pipeline damage, exposure, and abnormal activities such as anchoring or bottom-trawling near the pipeline.
• Search and emergency response. Locating people who have fallen overboard, missing targets, and objects on the seabed has traditionally required many vessels and long search times, making it costly and risky.
• Marine ecology and scientific observation. Individual identification of marine mammals such as cetaceans is difficult, and their diving behavior is irregular. Combining passive acoustic monitoring with unmanned platforms has become an important supplement for long-term observation.

Unlike the solo-submarine scenes in movies, real underwater search depends on sensors, algorithms, unmanned platforms, communication links, and cloud platforms. This article reviews four core questions from a systems perspective:

1. Can the system “see”? Multi-source sensing and signal processing.
2. Can the system “find” the target? Search coverage and route planning.
3. Can the system “keep track” of the target? Filtering, tracking, and multi-target data association.
4. Can the system “manage” the process? Data closed loops and cloud-based coordination.

2. Why the Underwater Environment Makes the Problem Difficult

Underwater search is difficult because the ocean differs fundamentally from land and air environments:

• Electromagnetic waves attenuate rapidly in water. GPS cannot be used directly underwater. Optical sensing usually has an effective range of only tens of meters, occasionally farther in clear water. Sound is the only practical means for long-range sensing. Acoustic sensing is therefore the foundation, with optical and magnetic methods used as near-field supplements.
• Multipath and reverberation create strong interference. Sound reflects repeatedly between the sea surface, the seabed, and other boundaries, producing multiple echo copies with different delays. The problem is especially severe in shallow water.
• The sound-speed profile changes with depth. Sound speed depends on temperature, salinity, and pressure, typically around 1450 to 1540 m/s. Field measurements are usually required because sound rays bend. Assuming a uniform sound speed can cause localization errors, and shadow zones can form near the thermocline.
• The channel is time-varying. Environmental noise comes from wind, waves, marine life, and shipping, and it changes with time and sea area. A fixed detection threshold is unlikely to remain effective for long.
• Platforms are constrained by energy and endurance. Unmanned underwater vehicles must balance propulsion energy, mission duration, and depth keeping. Real-time high-bandwidth transmission is often difficult to guarantee.

3. System Architecture: Five Cooperating Layers

An engineering-oriented underwater search and tracking system is usually organized into five layers:

Layer	Core responsibility	Key technologies
Sensing layer	Acquire acoustic, magnetic, optical, and other multi-source observations	Active sonar, passive arrays, side-scan sonar, multibeam echosounder, magnetometer, low-light camera
Edge computing layer	Real-time signal processing, detection, tracking, and data compression	Embedded signal processor, edge inference, track management
Transmission layer	Surface backhaul and near-field interaction	Underwater acoustic communication, RF link, satellite backhaul, wired export after recovery
Cloud platform layer	Storage, training, inference services, monitoring, and coordination	Object storage, message queue, container platform, logging service, monitoring and alerting
Application layer	Situational presentation, alert delivery, mission planning, and post-analysis	Spatiotemporal visualization, automatic report generation, open interfaces

The purpose of layering is to let each layer solve the problems it handles best. The edge focuses on millisecond-level real-time response. The cloud focuses on large-scale data processing and model iteration. Standardized data formats and message protocols connect the layers.

4. Sensing Layer: Combining Multi-Source Detection

4.1 Acoustic Sensing

• Active sonar transmits a known waveform and receives echoes. It provides long range and direct ranging, but it exposes the platform and may disturb marine life.
• Passive sonar only receives target radiated noise. Common features include shaft rate, blade rate, mechanical line spectra, and broadband continuous spectra. It does not expose the platform, but it depends on target noise and cannot detect silent targets.
• Side-scan sonar operates close to the seabed to produce high-resolution acoustic images. It is suitable for seabed objects, pipeline anomalies, and terrain surveys.
• Multibeam echosounder uses fan-shaped beam arrays to measure depths in multiple directions at once. It is suitable for large-area terrain mapping and anomaly localization.

A commonly overlooked engineering detail is that a target’s acoustic signature is strongly related to its operating condition. Propeller cavitation and mechanical noise change with speed and heading, so passive range estimates can drift. Algorithms must cover multiple operating conditions; otherwise, performance may drop sharply when a model is applied outside its training conditions.

4.2 Complementary Non-Acoustic Sensing

A magnetometer measures anomalies in the geomagnetic field. It is sensitive to ferromagnetic targets such as shipwrecks, pipelines, and anchor chains, and it is not affected by water turbidity. It is useful for detailed confirmation after acoustic detection identifies a suspicious area.

Optical cameras provide high resolution and help human analysts interpret target types, but they are limited by water clarity and lighting. They are better suited to close-range, point verification.

4.3 Combined Strategy

A mature approach is “coarse to fine, area to point.” First use side-scan sonar or magnetometers to screen a large area. Then send a vehicle or diver closer for optical and acoustic confirmation. Multi-source data must be aligned to a common spatiotemporal baseline, including timestamps and position coordinates, before fusion.

5. Signal Processing: Seeing Targets in Echoes

5.1 Matched Filtering and Beamforming

When active sonar uses a linear frequency-modulated pulse, matched filtering uses the time-bandwidth product to obtain processing gain, often written as the BT product, where B is bandwidth and T is pulse duration. It compresses a relatively long transmitted pulse into a narrow peak and improves signal-to-noise ratio. After compression, the main-peak width is approximately 1/B.

Beamforming adjusts the delay or phase of each array element and then sums the weighted outputs. This focuses received energy in a specific direction and provides array gain related to the number of elements. The introduction of spatial angle and time-delay information also provides a basis for time-difference localization.

5.2 Constant False Alarm Rate Detection

In real ocean environments, background variation can be severe. A fixed threshold either misses weak targets or produces many false alarms. CFAR estimates the local noise level around the detection point and adjusts the threshold dynamically. The following code is a simple illustrative implementation of cell-averaging CFAR:

def sliding_mean(x, radius):
    """Pointwise local mean; in engineering, optimize with prefix sums or integral images."""
    n = len(x)
    result = [0.0] * n
    for i in range(n):
        lo, hi = max(0, i - radius), min(n, i + radius + 1)
        result[i] = sum(x[lo:hi]) / (hi - lo)
    return result
def ca_cfar(sq, n_ref=16, n_guard=4, alpha=3.0):
    """
    Cell-averaging CFAR detection (illustrative implementation)
    sq     : squared envelope per cell (energy detection input)
    n_ref  : number of reference cells on each side
    n_guard: number of guard cells, used to prevent target leakage from raising the threshold
    alpha  : threshold multiplier, determined by false-alarm probability and background distribution
             (for example, exponential distribution after square-law detection)
    returns: list of cell indices that exceed the threshold
    """
    ref   = sliding_mean(sq, n_ref + n_guard)
    inner = sliding_mean(sq, n_guard)
    hits = []
    for i, e in enumerate(sq):
        if i < n_guard or i >= len(sq) - n_guard:
            continue
        # Estimate the reference-cell mean after excluding guard cells.
        noise_est = ((n_ref + n_guard) * ref[i] - n_guard * inner[i]) / n_ref
        if e > alpha * max(noise_est, 1e-12):
            hits.append(i)
    return hits

Engineering implementation also requires care:

• Clutter-edge problem. When a target lies between a strong reflector and a quiet background, the reference window may span two types of background, causing a sharp increase in false alarms. Order-statistics CFAR or parallel detection for multiple background types can be considered.
• Guard cells are necessary. Otherwise, target energy raises the threshold and causes self-masking.
• Energy detectors must guard against reverberation false targets. Rock outcrops and shipwrecks may have geometries similar to man-made targets, which is one reason to add multi-source fusion and tracking constraints.

5.3 Multistatic Localization

Multiple receiving nodes can receive the same transmitted signal synchronously, measure time difference of arrival, and solve equations to estimate target position. Long-baseline localization is an engineering application of this principle. It usually provides better accuracy than solutions based only on slant range. It requires synchronized clocks and good sensor geometry; poor geometry creates regions with weak accuracy.

6. Search Planning: How to “Find” the Target with Limited Range and Energy

6.1 Estimating Detection Time with Search Theory

The classic random-search model gives the probability of detecting a target after searching for time t, assuming no prior position information and a random target location:

P(t) = 1 − exp(−W · v · t / A)

where:

• W is sweep width, the effective lateral width in which a target can be detected during one pass when the target is present;
• v is platform speed;
• A is search area;
• t is elapsed search time.

This formula has two engineering implications. First, increasing coverage does not always reduce detection time. When W · v · t / A is about 1.6, detection probability is about 80%; at about 2.3, it is about 90%. Further increasing coverage yields diminishing returns. Second, the benefit of higher speed depends on detection performance. W is strongly related to signal-to-noise ratio and detection probability. Higher speed increases flow noise and self-noise, reducing effective detection range and therefore W. Propulsion energy grows roughly with the cube of speed, so increasing speed blindly can be counterproductive.

Sweep width W is determined jointly by detection-probability and sweep-width curves. These curves describe detection probability at a given lateral distance and the cumulative distribution of lateral distance. They require extensive field calibration and should not be assumed without evidence.

6.2 Standard Search Patterns

• Parallel-line search: parallel tracks spaced approximately W apart. It is suitable for a well-defined rectangular area. If the target direction is unknown, alternating headings, such as 90° and 0°, can reduce the chance of missing anisotropic targets.
• Expanding-square search: used when the initial target position is known and the target is essentially stationary. The search expands outward from the center, with each leg increasing by one leg length.
• Sector search: starts from the last known position and scans outward in sectors. It is suitable for fresh targets with the most recent position information.
• Intercept course: when target motion can be predicted, detection resources are placed along possible escape routes rather than near the current position.

6.3 Cost Functions and Algorithm Choices

Planning for underwater vehicles is essentially a constrained optimization problem:

• Objective: maximize expected detection probability or minimize expected detection time.
• Constraints: battery capacity and mission duration, vehicle maneuverability, obstacles and safety zones, collision avoidance, and degradation of positioning accuracy.

Common algorithm routes include:

1. Grid-based planning with A*-type algorithms: solves point-to-point paths, can incorporate obstacle costs, and is suitable as a basic planner.
2. Pheromone-style incremental graph search: converts unknown-area information gain into grid costs and is suitable for coverage-oriented search.
3. Sampling-based path planning: quickly finds feasible trajectories in high-dimensional state spaces and is suitable for platforms with maneuverability limits.
4. Sequential optimization and learning-based methods: can be trained extensively in simulation and used for online decisions, but they must be constrained by safety rules and human review. Models should not decide vehicle behavior without constraints.

7. Tracking Layer: Turning Points into Tracks

7.1 Single-Target Filtering

After detecting target plots, state-estimation methods are used to associate them into tracks. The state usually includes position, velocity, and optionally acceleration, or it can be simplified by separating two-dimensional horizontal motion from depth.

• Kalman filtering is the optimal recursive estimator under linear Gaussian conditions. It uses prediction and update steps.
• Extended Kalman filtering linearizes the nonlinear measurement equation using a first-order Taylor expansion at the current estimate.
• Unscented Kalman filtering propagates the nonlinear transformation through a set of deterministic sample points, called sigma points. It is usually more robust than EKF under weak nonlinearity.
• Particle filtering represents arbitrary distributions with many weighted particles. It is suitable for strongly nonlinear, non-Gaussian cases but requires more computation.

Typical underwater difficulties include:

• Inaccurate sound speed can cause range bias, so related uncertainty should be included in measurement noise.
• A single maneuvering platform may lack a velocity reference, and track accuracy can be affected by positioning-system errors.
• Measurement rate can fluctuate with communication and endurance limits, so the filter must tolerate intermittent track loss.

7.2 Multi-Target Data Association

When multiple targets appear at the same time, the core problem is deciding which plot belongs to which target. Common methods include:

• Global nearest neighbor: assigns plots by statistical distance, such as Mahalanobis distance. It is simple but prone to mismatch in dense clutter.
• Joint probabilistic data association: updates candidates that fall within the association gate using probability-weighted combinations.
• Multiple hypothesis tracking: retains multiple association hypotheses and delays decisions. It performs better with crossing targets and missed detections, but combinatorial growth requires pruning.

Track management is also essential. It includes track initiation, such as confirming a track when M out of N detections are hit, track maintenance, and track termination, such as when a track is lost for more than a threshold. Track quality scores can support alert grading and human-machine collaboration.

7.3 Multi-Platform Collaboration and Track Fusion

When multiple vehicles detect independently, time and space alignment must be performed first, followed by track fusion. Fusion methods such as covariance intersection and information filtering can avoid using the same measurement information repeatedly, which would otherwise overestimate confidence.

Collaboration depends not only on algorithms but also on underwater communication. Bandwidth is low, latency is high, and links may be interrupted. A practical strategy is local closed-loop operation, summary upload, and clarification on demand.

8. Cloud Closed Loop: Data, Training, and Engineering Governance

8.1 Data Layering

• Raw acoustic data: large in volume and impossible to regenerate. It is suitable for long-term retention in object storage, with hot and cold tiers based on access frequency to control cost.
• Processing results: detection plots, tracks, and alerts. These are small in volume and high in value density. They are suitable for message queues used by real-time consumers and downstream alerting.
• Metadata: mission, equipment, sea area, operating conditions, and calibration information. It supports scenario segmentation and performance attribution during training.

8.2 From Data to Models

A practical closed loop has five steps:

1. Collection: standardize formats, timestamps, and coordinate references.
2. Labeling: annotate plot-level detection boxes, target classes, and track IDs. Active learning can prioritize samples that the model finds uncertain.
3. Training: build reproducible container images, and record both data versions and model versions.
4. Deployment: export trained models into lightweight formats that can run at the edge. Quantization and pruning can reduce inference overhead.
5. Monitoring: track metrics such as detection rate, false-alarm rate, and track continuity. Bucket metrics by sea area and operating condition, and trigger retraining when degradation is found.

Degradation often comes from scenario drift. When moving to a new sea area, seabed sediment, noise background, and biological activity can all change. Dashboards organized by sea area and season make it possible to detect problems early.

8.3 Engineering Governance

• Logging service: centrally collect edge-device and cloud-service logs to trace where a missed detection occurred in the processing chain.
• Monitoring and alerting: set tiered alerts for device online rate, data-backhaul latency, and inference latency.
• Permissions and audit: control access to sea-area data and raw acoustic data by role, and keep operation records.
• Release process: use canary or gradual rollout for algorithm updates. Run a new version in a limited sea area first, compare performance, and then replace the old version fully after targets are met.

These generic cloud capabilities are commonly provided by major cloud platforms. Baidu AI Cloud also offers object storage, message services, container engines, logging services, and cloud monitoring, which can serve as one possible foundation for such data and operations systems. Service availability and scope should be confirmed for the target deployment region. Final selection should be based on actual metric testing and cost accounting.

9. Deployment Roadmap and Evaluation Metrics

9.1 Phased Deployment

1. Phase 1 — Observable: use a single platform with basic acoustic detection. Connect data collection, edge processing, cloud storage, and visualization, and accumulate real data.
2. Phase 2 — Discoverable: add multi-sensor fusion, CFAR optimization, and search planning. Validate detection metrics in typical sea areas.
3. Phase 3 — Trackable: add multi-target tracking, track management, and a collaborative platform to form persistent tracking capability.
4. Phase 4 — Replicable: platformize models, version algorithms, and systematize metrics, turning single-point capability into a repeatable engineering method.

9.2 Metrics

Metric	Meaning	Use
Detection probability	Proportion of true targets detected	Evaluate sensing capability
False-alarm rate / false-alarm density	Number of false alarms per unit time or unit area	Evaluate alert usability
Localization accuracy	Error between estimated and true position, such as CEP50 or CEP95	Evaluate whether secondary verification is needed
Track continuity	Number of ID switches and duration of track loss	Evaluate tracking stability
Detection time and energy consumption	Time from mission start to target detection and battery consumption	Evaluate planning and platform efficiency
Coverage	Effective scanned area as a proportion of planned area	Evaluate search execution quality

These metrics should be calculated by sea area and target type. Looking only at overall averages can hide serious degradation in specific operating conditions.

9.3 Human-Machine Collaboration and Compliance

• Confidence grading: high-confidence alerts can be pushed automatically; medium- and low-confidence alerts enter a human-review queue.
• Explainability: retain raw echo snippets and detection-chain snapshots for expert interpretation and error attribution.
• Marine environmental protection: active sonar effects on marine life are a general concern. Operations should follow the regulations applicable to the sea area and, when necessary, use passive monitoring or reduce transmit power.
• Data compliance: sea-area and pipeline information may involve sensitive data. Access permissions should be managed according to classification requirements.

10. Common Pitfalls and Practical Advice

1. Do not rely on a single sensor. Multi-source fusion is not optional; it is the foundation of underwater search.
2. Do not ignore sound-speed profiles and positioning baselines. Positioning errors often come first from missing environmental parameters rather than from insufficiently advanced algorithms.
3. Do not treat detection as tracking. A plot is not a track. Without track management, a system can be overwhelmed by false alarms.
4. Do not tune only in simulation. Simulation data cannot cover all details of real seabed sediment, biological noise, and platform disturbance. A real-data regression mechanism is necessary.
5. Do not pursue full automation without limits. When the cost of missed detection is high, a human-in-the-loop design is often more reliable than end-to-end automation.

Conclusion

“Undersea pursuit” may rely on dramatic tension in fiction, but in engineering it is a system method built from multi-source sensing, signal processing, search planning, filtering and tracking, and data closed loops.

For developers and enterprises, the real value is not a single algorithm. It is connecting sensing, decision-making, execution, and review into a measurable and iterable closed loop, so that every deployment and voyage becomes a basis for the next improvement. On cloud infrastructure such as Baidu AI Cloud, this closed loop can be built in phases and evolved step by step.