The Research Of Anti-occlusion Object Tracking Algorithm Based On Audio Auxiliary Information

Target tracking is an important topic in the field of human-computer interaction, but in the actual track, the target is susceptible to problems such as block. After the survey and analysis of the current research work, we conduct a study on the block from the two aspects of video track and audio-video track with the aim to improve the performance of tracking system under complex environment such as block. This paper has proposed a anti-occlusion fusion tracking method based on feature confidence, and improved model update mode. The main contributions of this thesis are as follows:1、In the video tracking, when the target is affected by light, shadow, and similar objects blocking, it is likely to cause the failure of video objuect tracking, a fusion algorithm based on feature confidence and similarity is proposed to solve this problem, it can take different fusion ways to different environments, and improve the system anti-jamming capability. The proposed algorithm uses the re-integration similarity to adjust the weights of the sum rule and product rule dynamically, while taking advantage of feature confidence to adjust each feature weight of the sum rule, to make the fusion results closer to the true state, and uses the confidence to detect occlusion, has overcome the problem that when the target is blocked by analogue, the similar feature has high similarity to not detect occlusion accurately. And it also has improved the update method of target model, when the occlusion does not occur, the model combines the information of initial template、previous frame template and the current model template, it can reflect the initial and current state to make the target template can adapt to the change of complex environment better; when the target is blocked, the proposed algorithm uses the previous frame template, which can be better to retain accurate information, reduce background noise integration, provide the basis for the target matching after block recovery, ensure the accuracy of tracking. Experimental results show that the proposed algorithm has better tracking performance in illumination、analogue occlusion and other environment, but also meets the requirements of real-time.2、Currently the single-mode tracking method is used more in speaker tracking. Audio track has a wide range of positioning、low computational complexity、good real time and other advantages, but is affected by background noise easily, and positioning accuracy is poor; Video tracking has high precision, but is susceptible to light、 shadow、occlusion and other complex environment, and the perspective is limited by the angle of the camera. Due to the good complementarity between audio information and video information, this paper presents a heterogeneous information fusion method based on feature confidence in the framework of particle filter to overcome the single-mode tracking defects, selects TDOA feature and color histogram feature, uses confidence to fuse the two features, giving full play to the mutual coordination between the two types of information, and improving tracking accuracy. As a result of a single color feature, so we use the confidence to assist the detection of occlusion to solve the problem of similarity occlusion detection error. Finally, we use the improved model update method to update the target template, give full play to the advantages of various types of information, ensure the accuracy and timeliness of template information. Experimental results show that the improved dual-mode tracking algorithm has better performance than single-mode tracking algorithm.