Lightweight multiobject ship tracking algorithm based on trajectory association and improved YOLOv7tiny

Abstract

In response to various challenges, such as high computational complexity, large parameter counts, and frequent vessel ID switching in ship tracking models, we propose a trajectory association and improved YOLOv7tiny-based vessel tracking algorithm, named the vessel tracking model with YOLOv7tiny-Vessel and StrongSORT-Vessel (VTM-YS). The algorithm consists of two stages: vessel detection and vessel tracking. In the vessel detection stage, we propose a lightweight Fast-PP-LCNet backbone network for efficient feature extraction of ships. Then, a novel Slim Bidirectional Pyramid Aggregation Network (Slim-BiPANet) is constructed to effectively fuse multiscale vessel features. Furthermore, introducing a coordinate attention (CA) module enhances the focus on vessel detection regions. In the vessel tracking stage, we introduce the OSNet reidentification network and the NSA Kalman filter algorithm into the StrongSORT model to reduce the frequency of tracked vessel ID switches. The results indicate that, in terms of vessel detection, the improved YOLOv7tiny achieves a 1.5% increase in mAP50 and a 13.9% reduction in parameters compared to the baseline YOLOv7tiny. For vessel tracking, the StrongSORT model achieves a 3.5% improvement in HOTA, a 7.0% improvement in MOTA, and a 73% reduction in ID switches, meeting the requirements for lightweight design and tracking accuracy.