Design And Implementation Of Single Target Tracking Algorithm Based On Deep Learning

Single target tracking is a very important branch of computer vision which is widely used in the field of intelligent video surveillance,human-computer interaction,visual navigation,virtual reality.It is usually faced with challenges such as scale change,shape change,background interference,illumination change,and motion blur in the single target tracking.In recent years,with the development of deep learning technology,the method based on deep learning has achieved good results in single target tracking,and also promoted the landing and application of single target tracking technology in all walks of life.Therefore,single target tracking technology based on deep learning has important research value.In this paper,through the research and analysis of Multi-Domain Convolutional Neural Networks,combined with its characteristics and shortcomings in the single target tracking,a new single target tracking algorithm ASTT-M is proposed.The algorithm realizes the anomaly detection and the multi-scale change of the target.A dual model regression strategy based on ridge regression and a similarity interference avoidance strategy based on search region cropping are proposed.Besides,a new way of sample update are designed based on clustering,and redefine the tracking control strategy.The target anomaly detection method is based on mean score of top5 candidate score anomaly detection.By this way,the anomaly which caused by target deformation,occlusion,scale change and illumination change can be detected in time during tracking.It will be processed in time when an abnormal situation occurs with tracking control strategy which is proposed in the paper.A multi-scale candidate box generation method based on region proposal network is proposed to fit the scale change of the target.We also adopts an idea of double model regression which considers the regression relationship between the initial feature,the current feature and location of the target in the regression of the initial tracking results.In order to avoid the influence of similar interference,a method based on search area clipping is used to keep the influence away from search area.Besides,this paper design a sample update strategy based on clustering which not only eliminates redundant samples and polluted samples,but also ensures the diversity and richness of the target samples.And a temporary sample set will be generated when an exception occurs.The online network not only keep a strong cognitive ability to the historical state of the target but also quickly learn the current state of the target with updating the network appropriately based on samples.In order to verify the effectiveness of the proposed method,we evaluated the results with the other nine classic single-target algorithms on the OTB50,OTB100,and OTB32 datasets.And then,we do comparison and analysis on accuracy and success rate.Besides,we select some video sequences with more difficult in tracking and compared with other trackers.Experiments show that the accuracy and success rate of the method on the test datasets are improved.