Abstract:

To solve the problem of constrains between the accuracy and speed for convolutional features for visual tracking methods, an algorithm namely adaptive position switching based on correlation filters was proposed. Pool4 and Conv5-3 layers were selected for features extraction. At the same time, effective features were obtained by the average feature energy ratio, which improved the tracking speed. Then it trained correlation filters with different Gaussian distributions of samples. Therefore, the best classifier was selected to predict the position according to the peak-side-lobe ratio, with a promotion in the generalization ability of the tracker. Finally, the sparse model update strategy was adopted to reduce the over-fitting and further speed up the algorithm. This algorithm was tested on OTB100 benchmark dataset. Tracking results demonstrated that the accuracy was 88.8%, 6.1% higher than the hierarchical convolutional features for visual tracking method. The tracking speed was 47.5 frames per second, which was 5 times than the original method, and showed favorable real-time performance.

Key words: visual tracking, adaptive features, correlation filter, peak-side-lobe ratio, model update, Gaussian distribution