IMM-Sonar-Sort: Multi-target Real-time Detection and Tracker of Underwater Pipelines based on Sonar Images

In this paper, a multi-beam forward-looking sonar image network framework is proposed for target detector and tracker in order to solve the problem of pipeline inspection and specific target tracking in complex underwater areas. The IMM-Sonar-Sort network framework uses supervised learning to detect objects in real time, and tracks the detected semantic information for multi-targeting. In the IMM-Sonar-Sort detection part, YOLOv10, an improved end-to-end real-time detection framework with variable serpentine convolution, is used to train and detect the weights of specific underwater complex datasets. In the tracking part, Interacting Multiple Model (IMM) and SORT (Simple Online and Realtime Tracking) algorithm is used to realize the real-time stable tracking of the maneuvering target by the sonar detector with the help of probability update and estimation fusion. The pool experiments show that the algorithm proposed has obvious effects on the detection and tracking of three typical characteristic targets: underwater underwater pipeline, underwater cable and underwater robot, and ensures the detection time (79ms), especially the target detection and tracking under the condition of vehicle variable speed.