Single Target Tracking Algorithm Based On Adaptive Fusion Of Multi-layer Convolution Features

Fast moving target and occlusion often result in poor tracking robustness and low tracking accuracy in complex video scenes. Aiming at such problem, this paper proposes a Siamese target tracking algorithm with adaptive fusion of multilayer convolutional feature tracking results. Firstly, the target is tracked using the Siamese tracking framework, which uses a wider and deeper backbone network, Resnet22. Then, the multi-layer convolution features of the frame image in the convolutional neural network are extracted and tracked. Finally, the forward tracking and backward tracking schemes are used to locate the target position by comparing the results of forward tracking and backward tracking, and the adaptive fusion prediction is performed on the hierarchical CNN feature tracking results. The algorithm solves the problem that the backbone network of Siamese tracker is shallow and the deep learning cannot be fully utilized, avoids the defect that the single-layer features of the network represent the target information incomprehensively, and enhances the generalization ability of the algorithm. The performance of the proposed algorithm is verified on the RGB public test set, and compared with the existing Siamese tracking algorithm. The experimental results show that the proposed algorithm performs better in accuracy and success rate, and has better robustness in complex cases.