Adaptive convolutional feature selection for real-time visual tracking
Xiong Changzhen,Che Manqiang,Wang Runling(Beijing Key Laboratory of Urban Intelligent Control,Beijing;College of Sciences,North China University of Technology,Beijing)
In order to reduce the redundancy of convolution feature for visual tracking and improve the real-time and robustness of the visual tracking algorithm, a real-time tracking algorithm based on adaptive convolutional features selection is proposed. The algorithm uses the feature mean ratio of the object region and the search region to evaluate the convolution operator. Firstly, the convolutional layer with the largest number of convolutional channels satisfying the feature mean ratio threshold is selected, and then the effective convolutional features of the selected convolutional layer is extracted to train the correlation filter classifier. Finally a sparse model updating strategy is adopted to improve the tracking speed. The algorithm is tested on OTB-100 standard dataset. The experimental results show that the average distance accuracy of this algorithm is 86.4%, 2.7％ higher than the original hierarchical convolutional features algorithm. the average tracking speed is 29.9 frames / sec, which is almost 3 times faster than before.