Int J Performability Eng ›› 2019, Vol. 15 ›› Issue (7): 1802-1812.

### Target Tracking Algorithm based on Context-Aware Deep Feature Compression

Ying Wanga, Aili Wanga,*, Ronghui Wangb, Haiyang Liua, and Yuji Iwahoric

1. a Higher Education Key Lab for Measuring and Control Technology and Instrumentations of Heilongjiang Harbin University of Science and Technology, Harbin, 150001, China
b Heilongjiang Province Public Security Department, Harbin, 150001, China
c Computer Science, Chubu University, Aichi, 487-8501, Japan
• Submitted on  ;
• Contact: * E-mail address: aili925@hrbust.edu.cn
• About author:Ying Wang is a master's student at Harbin University of Science and Technology. Her research interests include image superresolution and image classification.Aili Wang received her B.S. degree from the Department of Electronic and Communication Engineering and her M.S. degree and Ph.D. from the Department of Information and Communication Engineering at Harbin Institute of Technology in 2002, 2004, and 2008, respectively. She joined Harbin University of Science and Technology as an assistant in 2004 and became an associate professor in 2010. She was a visiting professor at Chubu University in 2014. Her research interests include image superresolution, image fusion, and object tracking.Ronghui Wang received his B.S. degree from the Department of Electronic and Communication Engineering at Harbin Institute of Technology in 2002.Haiyang Liu is a master's student at Harbin University of Science and Technology. Her research interests include object tracking.Yuji Iwahori received his B.S. degree from Nagoya Institute of Technology in 1983 and his M.S. degree and Ph.D. from Tokyo Institute of Technology in 1985 and 1988, respectively. He joined Nagoya Institute of Technology in 1988 and became a professor there in 2002. He joined Chubu University as a professor in 2004. He has also collaborated with UBC since 1991, IIT Guwahati since 2010, and Chulalongkorn University since 2014.

Abstract: The main focus of target tracking is robustness and efficiency. Because of challenges such as background clutter, occlusion, and rotation, high robustness and efficiency cannot be achieved simultaneously. A tracking framework of perceptual correlation filter is improved to achieve high-speed computation between real-time trackers. The main contribution to high-speed computing speed comes from improved depth feature compression, which is realized by combining content-aware features with multiple automatic encoders. In the pre-training stage, an automatic encoder is trained for each class separately. In order to obtain the feature map suitable for target tracking, the orthogonal loss function is added to the training stage and the fine-tuning self-encoder stage. Experiments show that the improved algorithm demonstrates great improvement in accuracy and speed.