Kernel-Based Relocation Siamese Network for Real-Time Visual Object Tracking

Abstract

Siamese networks have been paid more attention to video tracking due to its superiority in balance accuracy and speed. Based on the convolutional feature cross-correlation between the target template and the search region, trackers with Siamese network can search for the best result in the candidate box to get the tracking result. However, existing Siamese tracking algorithms are often affected by motion blurring, low resolution, distortion and other issues that blur search region in solving video object tracking problems. This paper presents a candidate box area generation method based on kernel density function to relocate the search region when track failed. Specifically, the tracker proposed in this paper fuses deep feature and color feature to generate candidate boxes from which more accurate tracking results can be obtained, moreover, the color feature is easily to calculate to reach real-time speed. Finally, by improving the candidate box generation algorithm, the problem of tracking missing due to fast motion, blurring and other factors is effectively solved with less time consuming.