Study On Particle Filter Tracking Based On Visual Saliency

As one of the most important branches in computer vision filed, moving object tracking has wide range of applications, such as video retrieval, scene surveillance, criminal behavior analysis and so on. However, because of the effect of some uncertainties of dynamic scene and target itself, the actual process of object tracking is very complex, a good tracking algorithm should deal well with the challenges of illumination change, target occlusion, pose variations and so on. To solve these problems, there appears many object tracking algorithms, and they have advantages and disadvantages. In recent years, visual saliency has received wide attention. The mechanism of visual saliency is a simulation of biological visual property, especially human visual property, it has the ability of selecting and remembering and it can exclude the interference and locate the target using the memory of target features. In the framework of particle filter tracking, this dissertation focuses on the simulation of human visual property and proposes a moving object tracking algorithm based on visual saliency. The two points of this dissertation are as follows:Firstly, Due to the problem that the robustness is not strong if using one single color feature in the tracking algorithm, we propose a visual tracking algorithm based on the fusion of the color feature and the saliency feature. As we all know, color feature has good resistance to certain occlusion and scale variation, but it is sensitive to illumination change, and when the target pose changes, color feature model and its template update mechanism can easily lead to template drifting. This dissertation studied some saliency detecting algorithms and use some good of them to tracking process. When there is illumination change or pose variation, we make full use of the invariance of the salient degree of the object area in the saliency map, avoiding the drifting phenomenon if using one single color feature.Secondly, Due to the phenomenon that the salient degree of target area is not high when using some "bottom-up" saliency detecting algorithms to get saliency map, we propose a saliency detecting algorithm based on the priori information of the target and apply it to the tracking process. Many existing detecting algorithms make use of the phenomenon that the target scale is big and the location of the target is near to the center of the picture. However, in the actual scene of tracking, the target environment is often very complex, the scale of the target is sometimes small and the location of the target is sometimes not near to the center of the picture, these all make the tracking task a challenge process. However, the location of the object is known in the first frame, the robustness of the tracking algorithm is depending on the application of the priori information of the target. Giving a simulation of the remembering mechanism of human visual property, this dissertation extract the target color histogram in HSV color space and put it back to the whole picture to get the saliency map. Then, when partial occlusion happens, the object can be located using the information of the part that not be occluded on the object.The experimental results show that, after putting saliency feature into the tracking process, the accuracy of the algorithm has improved, and our saliency detecting algorithm based on the priori information of the target performed well, it can handle some general occlusions in the tracing sequence test.