Research And Application Of Video Target Tracking Algorithm

With the rapid development of information?communication network?computer technology and microelectronic technology,target tracking technology has been widely used in the fields of video surveillance?human-computer interaction?intelligent navigation and medical assistance,which has important research value.The process of target tracking is as follows:firstly,the target of interest is tracked to learn the feature,then the basic model is established,and the target information is obtained after in the input video sequence.Finally,the target tracking is achieved.In the actual complex environment,the target attitude changes?scale changes?rapid movement and other problems like light?occlusion as well as similar interference cause huge challenges to the accuracy of the tracking results.The research and main work of this paper on the target tracking algorithm are as follows:(1)Based on the framework of particle filter,a combination of texture features and color features is proposed to improve the features in order to overcome the shortcomings of the particle filter based only on color features in dealing with similar color interferences.The texture features used are not basic LBP features,but the more stable HLBP features.What's more,adaptive weights are used to adjust the proportion of color features and texture features in the tracking results.The algorithm that combines the texture features of HLBP is more stable and more accurate than the algorithm that combines the texture features of LBP in tracking.(2)Aiming at the shortcomings that the features of machine learning algorithm can only learn the shallow features automatically,the convolution neural network is used to learn the features.The convolution neural network can learn the high-level semantic features of the object which makes the classification of objects more clear.Due to the large amount of computation in neural network,the computation speed may not meet the real-time requirements.Therefore,this paper takes offline video feature extraction combining online fine tuning of network parameters to construct online learning model of convolutional neural network to track in real-time.In addition,this paper proposes to adaptively adjust the learning rate according to the output confidence,through adaptive learning rate we can control the update speed of positive and negative samples and fine tune the network parameters of all connection layers to ensure the accuracy of tracking results.(3)The target detection algorithm is combined with the tracking algorithm to build a video surveillance system for the automatic detection and tracking of moving objects.The target detection module uses the combination of the edge detection and the frame difference method to improve the detection accuracy while the the CamShift algorithm is applied to the target tracking module.With the CamShift algorithm,the shape of the object adaptively adjust the rectangular box to achieve sustained and stable tracking effect.In our detection and tracking system,the trouble of manual box selection is avoided which improve the intelligence of the system.There are still many things that can be promoted in this paper:Feature fusion in the framework of particle filtering can add spatial features to increase feature diversity,while the online tracking algorithm based on a convolutional neural network can be combined with the TLD algorithm to improve the ability to resist long time occlusion.Last but no least,the video surveillance system needs further research on dynamic background conditions.