Vehicle tracking in satellite videos: improved SiamRPN++ incorporating bounding box constraint optimization and inertia mechanism

Aiming at the issues of low tracking accuracy and model drift in complex scenes, such as occlusion, for satellite video vehicle tracking, an improved SiamRPN++ tracking method for satellite video is proposed to achieve high-precision sustained vehicle tracking. Firstly, considering the tiny size of satellite video vehicles, the bounding box size constraint of SiamRPN++ is reduced to 5 × 5 pixels, so as to obtain a more accurate target bounding box, which can reduce similar targets and background interference, and improve tracking accuracy. Then, the maximum value (F_max) and average peak-to-correlation energy (APCE) of the classification score map are calculated to monitor the target state. Finally, when the target is monitored to be occluded, the inertial mechanism is employed to predict and correct the position of the occluded target to achieve continuous tracking. Experiment results based on the expanded XDU-BDSTU dataset show that the precision rate and success rate of the proposed method may reach 90.22% and 55.26%, respectively, which are 15.11% and 15.52% higher than those of the original SiamRPN++, respectively. The proposed method may enable continuous, high-precision tracking of satellite video vehicles at 52 FPS, with excellent real-time performance.