Robust Context-based Tracking Using Color Attributes

Visual tracking is an important study topic in computer vision, the accuracy of tracking directly affects the more advanced visual processing. With the development of computers and a variety of imaging equipment, the application fields of visual tracking are becoming more and more widespread, such as motion recognition, intelligent monitoring, video retrieval, human-computer interaction, traffic flow monitoring, intelligent vehicle navigation. Object tracking is to estimate the position, the outline, the trajectory and other status information of the target in successive video frames via deterministic methods or probabilistic inference methods and provides the basis for higher target behavior analysis. Due to the influence of target’s deformation and motion, occlusion, camera angle changes, illumination changes, background noise and other factors, robust object tracking face greater difficulties in the complex and changeful realistic scenario.In this paper, a fast and robust tracking algorithm is proposed by exploiting color attributes in the dense spatio-temporal context. First, the adaptive low-dimensional variant of color attributes and image intensity from the target and its surrounding region are used to model the object appearance. Second, a spatial context model between the target object and its local surrounding background is learned in the frequency domain online. Finally, the object location in the new frame is determined by maximizing the new confidence map. The complex conjugate of context prior model is used to update spatio-temporal context model.Moreover, we add a scale estimation approach into context-based tracking algorithm using color attributes. First, we construct the local context feature set based on the tracked location with several scales to get the context prior model. The object location and object scale is determined by comparing the maximum response of the confidence map of several scales. Finally, update the local context feature set and spatio-temporal context model.In order to test the performance of the algorithm, both quantitative and qualitative analysis are performed on the CVPR2013 tracking benchmark with one-pass evaluation, temporal robustness evaluation, spatial robustness evaluation and attribute-based evaluations. The success plots and precision plots are used to analyze the performance of the proposed tracker for each challenging factor. Experimental results show that the proposed approach outperforms state-of-the-art tracking methods when the target undergoes fast motion, pose variation, illumination change, heavy occlusions and background clutters.