Visual Tracking Based On Multimodal Information Fusion

Computer automatic vision is gradually developing towards industrial modernization,which makes the vision based on computer vision image closer to industrial modernization and people's daily life.And video target tracking is also widely used in intelligent video monitoring,aviation,military,human interaction and biomedical fields.Therefore,how to improve the accuracy,robustness and real-time performance of video target tracking is a frontier problem in the field of computer video target tracking,which is worth exploring and has very important theoretical and practical value.Using high-level features of convolutional neural network in video target tracking may lead to the decline of tracker accuracy,a weighted location information combination method(ROC)based on evidence reasoning rule(ER rule)is proposed.The algorithm is based on Siamese RPN,according to the different response of convolution features extracted by convolution neural network for the same video sequence in each convolution layer,multiple convolution features extracted by convolution neural network are used as video target tracking features to track the video,Finally,multiple location information results will be generated,and these results will be regarded as multiple pieces of evidence.All the evidences are assigned weight and confidence by a specific method,and then all the results are fused by evidence reasoning rules to get the final location information.Finally,the experimental results show that the proposed algorithm can improve the accuracy of video target tracking.In the process of video object tracking,the object is occluded,which leads to the degradation of real-time performance and robustness.In this paper,a decision-making oriented multi-mode target tracking algorithm fusion method(MIF)is proposed.The Tracking-Learning-Detection(TLD)algorithm and Kernel-Correlation-Filter(KCF)algorithm are selected as the integrated subject tracking algorithm to track the selected video target.Firstly,the histogram algorithm is used to detect whether the scene is occluded or not.Secondly,the improved Multi-process Markov Decision Process(MPMDP)is used to make decisions based on the time context information.Finally,according to the decision results,the TLD algorithm or KCF algorithm is selected to track the target when the target is occluded or not.Through experiments on video sequences,it is verified that the tracking model proposed in this paper has good performance in robustness and real-time performance.