Target Tracking Techniques for UAV Aerial Images

Aiming at the requirements of target tracking applications of UAV aerial images, the traditional target tracking methods have problems such as too strong dependence on the target motion model, significantly reduced accuracy in occlusion or interference, difficulty in judging tracking failure, and poor real-time performance, which are difficult to meet the practical application requirements. In this paper, a fast target tracking model based on Siamese neural network is designed to make the algorithm have high real-time performance. Combined with actual application requirements, two usage modes are designed, one is automatic target recognition then entering tracking, and the other is manually selecting targets then entering tracking. Based on the actual UAV aerial image data and the public infrared dataset, the average tracking success rate can reach more than 90%, and the test performance on the NVIDIA Jetson Xavier NX board can reach 22 frames per second, which can meet the real-time requirements.