Research On Target Tracking Algorithm Based On Probabilistic Graphical Model

Target tracking is one of the most important research directions in the field of computer vision and pattern recognition. It has been widely used in visual surveillance, human-computer interaction, military guidance, intelligent transportation. During the tracking, the robustness of algorithm will be affected by the changes of attitude of target, background and illumination, including occlusion. Among the changes, occlusion is a difficulty during the tracking. The probabilistic graphical model is the fusion of probability and graph theory and it has been adopted in the field of computer vision, such as target detection, recognition, classification and tracking. The focus of our research is the problem of occlusion based on probabilistic graphical model to improve the precision and stability when the target is in occlusion. In this paper, the block matching algorithm is the basic framework. The main contents are as follows:Firstly, we proposed a method to extract the salient feature regions from the target based on local feature points. The salient feature regions(blocks) based on Harris corners clustered through K-means algorithm make full use of the main feature information contained in the target region and eliminate redundant information.Secondly, in the view of the fusion of the tracking results of the sub-blocks, we proposed the target tracking algorithm based on the salient feature regions and markov random field. The markov random field is used to model the sub-blocks from the target based on the spatial information between different sub-blocks and the local information of each sub-block. The markov random field obtained through this way will be more precise. After modeling, the probabilistic inference algorithm is adopted to infer the more precise locations of sub-blocks on spatial domain based on the markov random field model. The experiments we performed proved that this algorithm can track the target with higher precision, especially when the target is in occlusion.Finally, we proposed a algorithm based on salient feature regions and conditional random field. This method will obtain conditional random field model based on the information of the sub-blocks, the spatial constraint information and motion information between adjacent frames. Then the model is used to obtain the probability value that this sub-block belongs to the target region. The probability value will be the weight for each sub-block to affect the location of the target. The experiments show that the precision of this algorithm is comparable to the algorithm based on markov random field and can also deal with occlusion, the speed of tracking can satisfy the real time request.