Real Time Target Tracking Scale Adaptive Based on LBP Operator and Nonlinear Meanshift
Zhenghua Shu, Guodong Liu, Zhihua Xie · 2017
The classical meanshift tracking algorithm is widely used because of its small computation and real-time performance. But the robustness of the classical meanshift algorithm is poor, the target location accuracy is not high, and the algorithm stability is not good. So, this paper proposes a nonlinear meanshift algorithm is a combination of color and texture histogram representation, this paper used the two value model of local LBP texture operator block thought based on improved, effectively extract the main target model of edge and corner to more refined target; in addition, through the effective use of information of the target candidate moment to solve the region, following the changes of the scale and direction of target motion in the problem. Three features of the algorithm proposed in this paper combining the complementary: accurate description of target color feature, texture feature, edge feature, first through the Bhattacharyya similarity coefficient between selected high similarity of the target histogram, and then choose two times through the log function. Through different scenes video tracking experiments show that, combined with characteristics of the classical meanshift algorithm and improved the representation method compared to target improvement algorithm in the complex scene tracking has higher stability and robustness. Finally, the real-time target tracking algorithm is applied to the video surveillance and traffic video surveillance in important places. The real-time target tracking algorithm can effectively track the vehicle target.