Research On Object Tracking Algorithm Based On Siamese Network

As one of the key technologies of computer vision,object tracking has important practical application value in video surveillance,human-computer interaction,automatic driving and other fields.The main task of object tracking is to give the position and size of the object in the first frame of the video,and predict the motion state of the object in the subsequent frames according to the context information of the video sequence,so as to obtain the complete motion trajectory of the object.In recent years,object tracking algorithm based on siamese network has been widely concerned for its advantages of fast tracking speed,high accuracy and end-toend off-line training model.It is one of the mainstream object tracking algorithms at present.Due to the complexity of tracking environment and the randomness of object motion state,the tracking drift of object tracking algorithm in siamese network is easy to occur when the object is shielded and out of view.To solve this problem,this paper proposes a siamese network object tracking algorithm based on feature matching(SiamFM).The algorithm uses maximum peak response and average peak correlation energy to discriminate tracking confidence.When the tracking confidence is high,the current frame is tracked accurately and the tracking result is output.On the other hand,for the video frame with low tracking confidence,the object feature matching tracking strategy is adopted to get the coarse positioning matching centroid point,and then the SiamRPN tracker is used for re-detection to get the precise position of the object.The accuracy and success rate of SiamFM are 0.889 and 0.673 on OTB100 dataset,and 0.556 and 0.297 on VOT2018 dataset,respectively.The results show that SiamFM effectively improves the tracking accuracy of the algorithm.The tracking drift problem of algorithm in complex scene is improved.In order to improve the tracking speed of SiamFM algorithm,a fast feature matching siamese network tracking algorithm based on motion vector(FSiamFM_MV)is proposed in this paper.The fixed grouping and motion vector parameter strategy were designed to classify and discriminate different video frames,SiamFM was used to track the object,and the tracking speed was improved while the tracking accuracy was guaranteed.The results show that the tracking speed of FSiamFM_MV is 25.2%higher than that of SiamFM,which proves the effectiveness of the improved strategy.