Object Tracking System Based On Multi-Cue Integration

Video object tracking is an important field of computer vision, it was used widely in security monitoring, vision-based human-computer interaction and military fields. The variety of information from the initial target can be extracted by visual object tracking, the location, shape and other information of the object in the current frame can be found automatically. After years of exploration, researchers have made a lot of new tracking methods, it can be divided into feature-based approach, non-online learning based methods, region-based methods and model-based approach. In recent years, video-based particle filter tracking method was more and more attention by researchers. With the traditional Kalman filter algorithm, particle filter algorithm can solve high-dimensional nonlinear dynamic system of non-Gaussian problems very well. However, the traditional particle filter tracking algorithm mostly use only a single cue for tracking. In the case of complex environmental, this kind of tracking system is not strong enough. What's more, the various features of the object may change along whit environmental factors, such as the color cue will change with the light of environmental. Single feature tracking system most use color feature for tracking, but most of it was extracted from HSV color space. In this way, brightness and saturation of the object was lost, which making the result is not accurate.To overcome the limitations of single-feature tracking system, this paper presents a multi-feature fusion for tracking system. Taking into account the color and edge cues do not change at the same time, we fuse these cues in the framework of particle filter to enhance the robustness of the system in complex environmental. In order to retain more information of the object, a three-dimensional histogram is used to describe the color cue in RGB color space. In the feature fusion process, an adaptive feature weight assignment algorithm is used. This algorithm can give a different weights to different cues,which can improve the robustness of the system. Finally, comparing experiment was shown that the proposed multi-feature fusion algorithm can be stable tracking the target object in the complex background.