Integration of low-cost sensing systems for autonomous vessel detection: leveraging acoustic and vision information

Abstract

The paper presents a novel framework for automatic classification and detection of waterborne vessels, tailored explicitly to integrate with low-cost, low-power off-the-shelf sensors and hardware. This framework demonstrates the practicality of incorporating affordable hardware and sensors into unmanned surface vehicles (USVs) to achieve dependable real-time surveillance and reconnaissance capabilities. This initiative marks a significant achievement as it is the first to successfully extract both auditory and visual signatures of bottom trawling vessels, presenting compelling evidence to identify vessels engaged in the detrimental practice. The acoustic signal classification model utilizes the Mel Frequency Cepstral Coefficients (MFCCs) and employs a multi-class neural network model for accurate classification. The proposed model achieves an impressive testing accuracy of 95.42%, highlighting the effectiveness of MFCCs in clustering underwater acoustic signals. The visual component of the system utilizes the YOLOv3-tiny model and is optimized to facilitate faster inferencing. It is seamlessly integrated with the DeepSORT tracking algorithm, enhancing the overall detection capabilities. By combining the strengths of visual and acoustic subsystems, this integrated approach overcomes the limitations of each component individually. It provides a powerful solution for the detection of vessels and activities while offering a practical approach to maritime defence and ocean conservation