Vehicle Object Tracking Based on Feature Fusion Siamese Network Under Unstructured Terrain for Unmanned Ground Vehicle

Vehicle target tracking technology is an important research component for unmanned ground vehicle. It has wide applications in areas such as cooperative operations and battlefield surveillance. In the actual complex unstructured terrain, the object appearance is susceptible to interference such as light changes, similar backgrounds, and deformations, which bring great challenges for vehicle target tracking. In recent years, deep learning techniques have shown potential in solving target tracking problems under interference conditions due to their powerful feature representation capabilities. To solve the problem, a vehicle target tracking method based on Feature Fusion Siamese Network (FFSN) is proposed. It designs a two-stream feature extraction network. Low-level visual features and high-level semantic features are integrated together, which can distinguish deep features and improve its resolution. The experimental results demonstrate that the method can achieve accurate tracking under complex unstructured terrain in real time, the tracking speed under the GPU is 111 fps. Compared with the baseline GOTURN algorithm, the overlap accuracy and distance accuracy of the proposed method are improved by 8.17% and 13.16%.