The Research Of Object Tracking Algorithm Based On Siamese Neural Networks

With the continuous development of "Intelligent City",Object Tracking task,as one of the important fields of computer vision,has played an important role in all aspects of civil and military fields and has huge development space.However,due to the high requirements of Object Tracking on accuracy and timeliness,the combination of Object Tracking and deep learning has not yet achieved the best effect.As a new force emerging in the field of Object Tracking,twin networks have great development potential in tracking accuracy and tracking speed.Firstly,this paper introduces the basic theory of twin network and Object Tracking,and carries out the reproduction and test of SiamFC Object Tracking algorithm,the ice-breaking work of twin network in the field of Object Tracking.Through the analysis of the experimental results,four problems of SiamFC algorithm are proposed in this paper.(2)during target deformation,SiamFC tends to generate tracking drift;(3)the features extracted by SiamFC are not sufficiently representational;(4)SiamFC does not deal with redundancy of feature channels.Aiming at the above four problems,this paper proposes two tracking algorithms,IeSiam and StSiam,and proposes solutions from the aspects of model off-line training strategy and online tracking respectively:(1)since SiamFC adopts a completely offline training mode,designing a better offline training strategy will improve the performance of the algorithm.IeSiam algorithm is based on the Ie-loss model training strategy proposed in this paper to solve the problem of insufficient processing of SiamFC offline training.The Ie-loss training strategy includes the following three aspects:the Ie-loss loss function is designed to eliminate the unbalanced distribution of positive and negative samples.The difficult sample data enhancement strategy is designed to increase the training samples with similar interference,randomly change the contrast,brightness and displacement of the samples,and improve the anti-interference of the network.Large marker data sets in the target detection domain,Youtube-bb and COCO data sets,were selected for model migration.In order to verify the effectiveness of Ie-loss training strategy,IeSiam algorithm adopts the same network structure as SiamFC algorithm and adopts Ie-loss model training strategy for training.Experimental results show that the accuracy and success rate of IeSiam algorithm have been improved,and the performance of the algorithm has been significantly improved in the case of target deformation and similar background interference.(2)StSiam algorithm designed a solution to improve the semantic representation ability of features,aiming at the tracking drift problem and continuous tracking failure caused by a single feature description in the online tracking process of SiamFC algorithm,and obtained superior tracking performance.StSiam algorithm includes the following three aspects:for the problem of weak representativeness of features,the depth convolution features of AlexNet are merged.To solve the problem of feature channel redundancy,adaptive channel selection module is designed,including channel screening based on target image and channel selection strategy based on gradient.Aiming at the problem of target deformation,the anti-interference capability of tracking algorithm is improved,and the template updating mechanism is introduced.Finally,the StSiam algorithm was tested on the otb-100 dataset and the test results were analyzed.StSiam achieved good performance and effectively eliminated the ambiguity of the tracking process.To sum up,this paper studies and analyzes the application of the twin network in the Object Tracking task from two aspects of model off-line training strategy and online tracking process,and experiments and deduces the proposed theory and method on the recognized data set,and obtains good experimental results.