Combining Denoising Autoencoders and Dynamic Programming for Acoustic Detection and Tracking of Underwater Moving Targets

Alberto Testolin, Roee DiamantSensors 2020 · DOI

TL;DR

This paper introduces a novel method, CDA-TBD, that combines deep learning (Convolutional Denoising Autoencoders) with dynamic programming (Viterbi algorithm) to accurately detect and track underwater moving targets using sonar. This approach significantly improves performance in noisy environments with low signal-to-clutter ratios, outperforming existing methods.

What Problem Does This Paper Address?

Detecting and tracking moving targets underwater using sonar (SONAR) is a significant challenge due to the highly stochastic nature of acoustic measurements, often obscured by noise. This is crucial for applications ranging from marine ecology research (e.g., estimating fish biomass) to security. Existing methods, like track-before-detect (TBD), struggle with computational complexity and accurately modeling the often unpredictable motion of marine animals. Furthermore, traditional sonar systems often require large arrays, which are complex and expensive to deploy. The paper aims to address these limitations by developing an efficient, single-transceiver system capable of real-time detection and tracking in challenging, noisy underwater conditions.

Methods

The core of the proposed method, CDA-TBD, involves two main stages:

  1. Convolutional Denoising Autoencoder (CDA): This deep learning component acts as a sophisticated noise filter. It takes a time-distance (TD) matrix, which represents sonar reflections over time and distance, as input. The CDA is trained to remove clutter and highlight potential target reflections, which often appear as curved lines in the TD matrix. This pre-processing step is crucial for improving the signal-to-clutter ratio (SCR).
  2. Dynamic Programming (Viterbi Algorithm): The denoised TD matrix from the CDA is then fed into a modified Viterbi algorithm. This algorithm, a form of dynamic programming, is used for probabilistic tracking. It identifies sequences of potential target detections (curved lines) within the denoised matrix and determines the most likely path of a moving target. The method is adapted to handle multiple targets and is optimized for computational efficiency.

The paper also details a data augmentation strategy to create a large, statistically representative dataset for training the CDA, which is a critical aspect for deep learning models.

Datasets

The paper utilizes a custom dataset collected from over 50 sea experiments conducted in various marine environments. This dataset includes:

  • Approximately 1000 target-based TD matrices: These matrices contain reflections from identified targets (e.g., divers, fish, sharks).
  • Over 20,000 clutter-based TD matrices: These matrices consist solely of noise and stationary reflections.

The dataset was augmented to create a larger training set of 10,000 target-based TD matrices. The authors emphasize the importance of using real-world data due to the difficulty in accurately simulating underwater acoustic channels and target reflections. This dataset is made freely available for research.

Key Insights and Ideas

  • Deep Learning for Noise Reduction in Sonar: The paper demonstrates the power of Convolutional Denoising Autoencoders in effectively filtering out significant noise from sonar data, a critical step for improving target detection.
  • Combining Deep Learning and Traditional Algorithms: The synergistic combination of CDA for denoising and the Viterbi algorithm for tracking proves highly effective, leveraging the strengths of both approaches.
  • Track-Before-Detect (TBD) with Enhanced Efficiency: By using a denoised TD matrix, the computational complexity of the Viterbi algorithm is significantly reduced, making real-time TBD feasible for underwater applications.
  • Importance of Real-World Data: The authors highlight that simulating underwater acoustic environments is challenging, underscoring the value of their large, experimentally collected dataset for training and validating their model.
  • Single Transceiver Viability: The success of the method with a single omni-directional transceiver opens doors for more cost-effective and deployable marine monitoring systems.

Key Results

  • Improved Detection and Tracking: The CDA-TBD method significantly outperforms benchmark methods (CDA-Max and Viterbi) in both detection and tracking accuracy, especially in low Signal-to-Clutter Ratio (SCR) conditions.
  • Favorable Trade-off: The system achieves a favorable trade-off between detection and false alarm rates, as evidenced by Receiver Operating Characteristic (ROC) curves.
  • High Accuracy in Low SCR: The method demonstrates robust performance even at SCRs as low as 4 dB, with tracking errors achieving sub-meter accuracy at 8 dB SCR.
  • Scalability to Multiple Targets: The approach is shown to effectively handle scenarios with multiple moving targets.
  • Computational Efficiency: The CDA-TBD method offers a substantial reduction in computational complexity compared to traditional TBD approaches.

Limitations and Future Work

  • Threshold Setting: The paper notes that setting the detection threshold (T₁) for the CDA-TBD approach might require training, whereas other methods like CDA-Max allow threshold setting based on prior knowledge of target characteristics.
  • Unsupervised Learning: The authors suggest exploring unsupervised learning for CDA training as a future research direction, potentially using anomaly detection schemes.
  • Real-time Implementation: A key future goal is to implement a real-time version of the CDA-TBD method for their SYMBIOSIS monitoring platform.
  • Extended Range: Future work will also focus on extending the operating range of the detection system.

Summary updated

BibTeX
@article{testolin2020combining,
  author = {Alberto Testolin and Roee Diamant},
  title = {{Combining Denoising Autoencoders and Dynamic Programming for Acoustic Detection and Tracking of Underwater Moving Targets}},
  journal = {Sensors},
  year = {2020},
  doi = {10.3390/s20102945},
}
Figure