The Research Of Single Speaker Tracking Algorithm Based On Video Multi-feature Fusion In Meeting Room Environment

With the constant development of internet of thing technology, computer vision and the continual advancement of image processing algorithm, speaker tracking by visual multi-feature fusion method and location has become a significant research project in current tracking and location field. A typical application is the speaker tracking and location in smart meeting room environment. So this thesis is mainly to study and discuss the problem of multi-feature fusion in the object tracking, the main works are as follows:(1) Considering defects of Dempster’s combination rule in the treatment of highly conflict evidence, two improved method are proposed from the perspective of correction evidence source. One is based on the batch of correction evidence source, Joint similarity of each evidence were obtained by introducing the cosine similarity function and the distance of the bodies of evidence, according to joint similarity getting weighted coefficient of each evidence. The Dempster combination rule was adopted to realize information fusion after evidence weighted average in the system. Comparing with other methods, this method not only considered the distance similarity, but also considered the cosine similarity, the proposed approach is more reasonable than other methods. The other method is the part of evidence source correction based on the conflict detection, Firstly, the mutual support degree of the every evidence is obtained by evidence distance, and the support degree is normalized to the credibility degree of evidence.a "referenced evidence" is obtained by averaging all evidences according to support degree. And then the referenced evidence is used for deviation verification and modification of the original evidences.Lastly, the Dempster’s rule of combination was adopted to combined evidences after disposing. The results of numerical examples show that this new method improves the reliability and rationality of the evidence combination results.(2) For the evidence theory can not effectively distribute conflict information, a new combaiton rule based on local conflict distribution strategy is proposed. According to evidence distance, the credibility of each evidence and the weighted average evidence were obtained. Then every focal elements’absolute distance was got, according to absolute distance the credibility of each focal element were obtained, which used as local conflicts allocation factor. Finally, the results of numerical examples show that the proposed method improves the reliability and rationality of the evidence combination results.(3) For the traditional conflict coefficient can not effectively represent the relationship between the evidences, Evidence adaptive combination algorithm based on the new measure criteria of evidence conflict is proposed, a new representation method of conflict evidence combined classical conflict coefficient is defined by linking pignistic transformation, experimental results showed that:a new conflict factor can effectively represent the relationship between the evidence.(4) For the problem of the common fusion stratey:product rule and weighted sum rule can’t effectively fuse the video multi-feature, the three improved method are proposed. The first method is that the improved evidence theory is used to fuse video multiple features in target tracking system to finish complete an accurate tracking. The K-L distance and the uncertainty of each evidence are used to obtain the weights of each evidence to correct the combined evidence, and finally the improved method is used to fuse video multiple features in target tracking system. The simulation results show that the proposed method has better performance. The second method is multi-feature adaptively fusion based on feature distance which intergrate product rule and weighted sum rule into the fusion framework. According to feature distance adaptively adjust the weights of product rule and weighted sum rule. Simulation results show that:the fusion method can achieve the goal of accurate tracking. The third method is the adaptive fusion strategy based on loss coefficient which redistribute the the loss information. Simulation results show that the method has the advantages of product rule and weighted sum rule, and can improve the accuracy and robustness of the tracking process.