A lightweight multi-target ship tracking model based on Yolov7

Abstract

Abstract Ship multi-target tracking has long been a topic that has attracted researchers from different fields. Currently, the following challenges still exist in ship multi-target tracking. Ships obscuring each other leads to an increase in misdetections and omissions in the algorithm. Some algorithms have a large number of parameters and computations, which are not favourable for deployment into devices. As a result, we present a novel multi-target tracking technique that combines the Yolov7 detector with Kalman filtering. First, the retrieval of fine ship details in the video is accomplished by employing the CNNS(Convolutional Neural Networks) + Transformer + CNNS architecture. Second, we introduce a novel lightweight module known as Light-SPP, which aims to integrate ship features. Finally, Wise-iou Loss, which increases the predictability of ship position, is shown as the detector’s localization loss function. The experimental results show that the number of parameters and computation of the model decrease by 11.0% and 17.7% respectively, and the continuous tracking accuracy and tracking and positioning accuracy improve by 17.7% and 7.4% respectively. In summary, our proposed ship multi-target tracking algorithm mainly solves the problems of low tracking accuracy and large computational volume, and has high engineering application prospects in the field of water transport.