💻 computer science

Multi-AUV Marine Life Tracking with Single Hydrophone Payloads via a Hidden Markov Model Equipped Particle Filter

This paper presents a multi-AUV marine life tracking system that utilizes single hydrophone payloads and a Hidden Markov Model-equipped particle filter to achieve high-resolution acoustic localization with significantly lower drag and improved accuracy compared to traditional multi-hydrophone AUV systems.

Original authors: Christopher Herrera, Kehlani Fay, Christopher Clark, Alberto Soto, Christopher Lowe, Mario Espinoza

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Christopher Herrera, Kehlani Fay, Christopher Clark, Alberto Soto, Christopher Lowe, Mario Espinoza

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to find a specific friend in a massive, foggy swimming pool. Your friend is wearing a special whistle that blows a short, distinct "toot" every 8 seconds. You can't see them, but you have a super-sensitive ear that can hear the whistle.

This paper is about a team of researchers who figured out how to track these "whistling" marine animals (like sharks) using a fleet of underwater robots (AUVs) that are much smaller, faster, and cheaper than previous versions.

Here is the breakdown of their work using simple analogies:

The Problem: The "Heavy Backpack" Issue

In the past, to find a whale or shark, scientists used underwater robots that carried two microphones (hydrophones) on a big, heavy frame.

The Analogy: Imagine trying to run a marathon while wearing a heavy backpack with two giant speakers on it. It's hard to move fast, and the extra weight drains your energy (battery) quickly.
The Consequence: Because the robots were slow and heavy, they couldn't keep up with fast-moving animals. Also, if the animal swam out of range of a single stationary microphone, the data was lost.

The Solution: The "Swarm of Lightweights"

The researchers came up with a new strategy: instead of one big robot with two microphones, they used multiple small robots, each carrying just one tiny, lightweight microphone.

The Analogy: Instead of one slow runner with a heavy backpack, imagine a team of five sprinters, each carrying a single, feather-light ear. They can move much faster and cover more ground together.
The Benefit: These robots are sleek and fast, allowing them to chase highly active animals without getting tired or creating too much water resistance (drag).

The Magic Brain: The "Detective Algorithm"

The tricky part is that one microphone can only tell you how far away the sound is, not which direction it's coming from. It's like hearing a car honk and knowing it's 100 meters away, but not knowing if it's to your left or right.

To solve this, the team used a computer program called a Particle Filter equipped with a Hidden Markov Model (HMM).

The Particle Filter (The "Guessing Game"): Imagine throwing thousands of tiny, invisible darts into the water, all guessing where the shark might be. As the robots hear the whistle, the program checks the darts. If a dart is in a spot where the sound should have been heard, it stays. If a dart is in a spot where the sound shouldn't have been heard, it gets removed. Over time, the darts cluster together around the shark's true location.
The Hidden Markov Model (The "Behavioral Detective"): This is the smart part. The program doesn't just guess randomly; it learns how sharks actually move. It knows that sharks don't usually teleport or spin in circles instantly. By studying past shark movements, the program predicts the most likely path the shark will take next.
- Analogy: If you are tracking a dog, a random guess might think the dog could be anywhere. But a "behavioral detective" knows the dog is likely to chase a ball or follow a scent, not suddenly appear on the other side of the city. This helps the robots find the shark faster and more accurately.

The Results: How Well Did It Work?

The team tested this system in two ways:

Real Ocean Tests: They tracked a stationary transmitter (acting like a shark) off the coast of Long Beach, California, and a moving one in Costa Rica.
- Result: For short trips, the system was accurate to within about 10 meters (roughly the length of a bus).
Computer Simulations: They ran a massive virtual simulation using real data from 22 white sharks tracked over months.
- Result: Even over long distances and large areas, the system stayed accurate to within about 15 meters.

Why the "Detective" Was Better

The researchers compared their "Behavioral Detective" (HMM) against two simpler guessing methods:

The "Random Walker": Assumes the animal moves in a completely random direction. (This failed often).
The "Constant Speed" Walker: Assumes the animal moves at a steady speed in a straight line. (This was okay, but not great).
The "Behavioral Detective" (HMM): This one won. Because it understood the habits of the sharks, it could figure out the shark's location even when the robots lost the signal for a while or when the data was messy.

The Bottom Line

This paper proves that you don't need a giant, expensive, slow robot with a heavy microphone setup to track marine life. By using a swarm of small, fast robots and a smart algorithm that understands animal behavior, scientists can track animals more accurately, for longer periods, and with less disturbance to the ocean environment.

The system is robust enough to handle losing up to 20% of the sound signals without failing, and even with up to 80% signal loss, it can still eventually find the animal, though it takes a bit longer to figure it out.

Technical Summary: Multi-AUV Marine Life Tracking with Single Hydrophone Payloads

Problem Statement
Tracking marine animals is essential for understanding migration, behavior, and the impacts of climate change. While acoustic telemetry allows for continuous data collection via transmitters, traditional stationary hydrophone arrays suffer from high data loss when animals move out of range. Conversely, active tracking using surface boats is labor-intensive and offers limited spatial resolution. Previous Autonomous Underwater Vehicle (AUV) solutions attempted to localize transmitters using multi-hydrophone payloads mounted on large frames. While effective for angle-of-arrival estimation, these heavy payloads induce significant drag, reducing AUV speed, maneuverability, and battery life, thereby limiting their ability to track highly mobile animals. Furthermore, alternative solutions using transponding transmitters require custom, large transmitters that are only feasible for large species like sharks.

Methodology
This work proposes a system utilizing multiple AUVs, each equipped with a single, compact, omnidirectional hydrophone, to localize an acoustic transmitter. The system operates under the assumption that the transmitter and AUVs remain near the surface, allowing for 2D state estimation.

Distance Estimation: Since the Time of Flight (TOF) cannot be measured directly without synchronized clocks, the system estimates distance changes ( $\Delta r$ ) by analyzing the difference between consecutive signal detection intervals ( $\Delta t$ ) and the transmitter's known pulse period ( $T$ ). The change in TOF is calculated as $\Delta TOF = \Delta t - T$ . By accumulating these changes from a known initial distance, the system estimates the absolute distance between the AUV and the transmitter.
State Estimation via Particle Filter: A particle filter fuses measurements from multiple AUVs to estimate the transmitter's state vector: position ( $x, y$ $x, y$ ), orientation ( $\theta$ $θ$ ), linear speed ( $v$ $v$ ), and angular velocity ( $\omega$ $ω$ ).
- Prediction Step: Particles are propagated forward in time using one of three motion models.
- Correction Step: Particle weights are updated based on the likelihood of the observed distance and distance rate of change ( $\dot{r}$ ) given the particle's predicted state.
- Resampling: Particles are resampled to focus computational resources on high-probability regions.
Motion Models: The study compares three distinct motion models:
- Baseline Random Walk: Adds Gaussian noise to coordinates.
- Velocity Model: Adds Gaussian noise to linear and angular velocities.
- Hidden Markov Model (HMM): A novel approach where the motion model is derived from historical white shark trajectory data. The HMM assumes the animal switches between unobservable behavioral states, with observable velocities distributed according to state-dependent Gamma (linear speed) and von Mises (angular speed) distributions.

Key Contributions
The paper outlines five primary contributions:

A method to measure distance to off-the-shelf acoustic transmitters using a single omnidirectional hydrophone on an AUV.
A novel multi-AUV localization algorithm employing a particle filter with an HMM-based motion model.
Offline validation of the algorithm using real-world ocean deployment data and extended simulation trials.
A comparative analysis demonstrating that the HMM motion model outperforms both standard velocity and random walk models.
An analysis of the system's convergence and robustness under varying levels of measurement loss.

Results
The system was validated using three datasets: a Long Beach dataset (stationary transmitter, moving AUV), a Santa Elena Bay dataset (moving transmitter and AUVs), and a large simulated dataset based on 124,000 white shark position observations.

Accuracy:
- Short-term/Real-world: The system achieved a Root Mean Square Error (RMSE) of approximately 10 meters for short-term deployments.
- Long-term/Simulated: For longer trials over larger areas, the RMSE was approximately 15 meters.
Model Performance:
- In the Long Beach dataset (simple movement), all models performed similarly (RMSE ~4.4m to 6.1m).
- In the Santa Elena Bay dataset (complex movement), the HMM model achieved the lowest RMSE (8.78 m), outperforming the velocity model (12.31 m) and baseline (11.22 m).
- In the Simulated dataset (large scale), the HMM model significantly outperformed others with an RMSE of 13.62 m, compared to 30.77 m for the velocity model and 970.46 m for the baseline. The HMM was particularly effective at maintaining multi-modal distributions (handling ambiguity between two possible intersection points) without diverging.
Convergence Time: The HMM model generally took longer to converge (average 180 s in Santa Elena Bay) compared to the baseline (27 s) and velocity (40 s) models, as it maintained particle diversity across ambiguous locations longer before resolving the correct state.
Robustness: The system remained robust with up to 20% measurement loss (average convergence time ~44 s, RMSE ~15 m). Performance degraded gracefully up to 80% loss, though convergence times and RMSE increased significantly. Above 80% loss, performance deteriorated drastically.
Depth Assumption: The 2D assumption was validated against shark depth data, showing that for depths up to 10.7 meters, the error in 2D distance estimation is negligible at typical tracking ranges (e.g., <2.4% error at 50m).

Significance
The paper claims that this work provides a practical alternative to heavy, multi-hydrophone AUV payloads. By equipping multiple AUVs with single, compact hydrophones, the system reduces drag, allowing for higher speeds and better maneuverability to track highly mobile marine life. The integration of an HMM motion model, fitted to historical animal behavior, improves localization accuracy over generic motion models, particularly in complex, real-world scenarios. The system demonstrates that accurate state estimation (approx. 10–15 m RMSE) is achievable even with significant data loss and without requiring specialized, large transmitters. This approach enhances the efficiency of data collection and offers a "stealthier" method for tracking that minimizes stress on the environment and the animals, ultimately aiding in the identification of areas requiring conservation protection.