Hierarchical Convolution Feature for Target Tracking with Kernel-Correlation Filtering

Abstract

Target tracking is widely used in many fields, but tracking performance still needs to be improved due to factors such as deformation, illumination and occlusion. In this paper, we propose a scale adaptive target tracking solution based on hierarchical convolution features and establish a kernel correlation filtering target tracking framework that combines multi-layer convolution features. The improved convolutional neural network is used to extract multi-layer features, and the correlation filters of each layer are separately trained to perform weighted fusion to obtain the target position. Then, the edge box algorithm is adopted to obtain the size of the actual tracking frame to achieve exact target tracking. An extensive evaluation on OTB-2013 with public test sequences are conducted. Experimental results and analysis indicate that our method is better than other known advanced tracking algorithms even in video sequences with many uncertain factors, while the speed and accuracy of tracking can be effectively improved.