A research of target tracking algorithm based on deep learning and kernel correlation filter
Shengtao Sun, Jibing Gong, Yangyang Li, Lizhe Wang, Kaisheng Wang · 2022 International Joint Conference on Neural Networks (IJCNN) · 2022
As a hot topic in computer vision, target tracking has a vital application in many scientific and technological fields. The tracking method based on correlation filtering transforms the target tracking from the time domain to the frequency domain through the Fourier transform, which can boost the accuracy and success rate. However, in the complex tracking environments, the target tracking process may be affected by deformation, occlusion, and other inferences, which make the traditional target tracking algorithms hardly accommodate the requirements of robustness. Aiming the target tracking in complex scenes, this paper tries to improve the feature extraction based on the Convolution Neural Network, which can learn deep features of the target from different convolution layers with more abstract characteristics. Then these multiple features are fused to enhance the robustness performance of the traditional Kernel Correlation Filter algorithm from the aspects of model characteristics. Furthermore, the accuracy and success rate of the proposed algorithm are verified based on comprehensive comparative experiments in the Object Tracking Benchmark with variant interferences.