Object Tracking Based On Improved Siamese Network

Visual object tracking technology is a basic research topic in computer vision and video analysis.Its task is to locate the center position of the object and estimate the size of the object boundary box in the changing video sequence.It is widely used in intelligent video surveillance,human-computer interaction,robotics,automatic driving and other scenes.Although the object tracking technology has made great progress in recent years,the performance of the existing object tracking algorithm still needs to be further improved to meet the practical application requirements due to the existence of many challenges such as occlusion,background clutter,illumination change,scale change,motion blur,fast motion and deformation.(1)The siamese neural networks object tracking algorithm integrating distractor-aware model is proposed in consideration of the problem that the fully-convolutional siamese networks algorithm for object tracking(SiamFC)is prone to tracking failure in the case of heavy occlusion,rotation,illumination variation,scale variation,etc.Firstly,the low-layer structural feature and the high-layer semantic feature are extracted from siamese networks and they are fused effectively to improve the representation ability of the feature;Secondly,the template adaptive strategy is used to update the template online,so that the tracking accuracy is improved in the case of occlusion and rotation.At the same time,the distractor-aware model based on color histogram features is introduced into the algorithm.The object response map is obtained by weighted fusion to estimate the position of the object and the adjacent frame scale adaptive strategy is used to estimate the optimal scale;Finally,the 2015 object tracking standard dataset(OTB100)was used to test the performance of the SiamDam algorithm.The experimental results show that the overall tracking accuracy and the overall success rate of SiamDam algorithm are 0.851 and 0.800,which are 10.3% and 9.6% higher than the SiamFC algorithm,respectively.(2)The long-time object tracking algorithm(SiamCsLT)via co-occurrence statistics and siamese network is proposed consideration of the problem that the SiamDam algorithm may fail to track during the tracking process and cannot be applied to long-term object tracking scenarios.Firstly,the object appearance similarity,peak-to-side lobe ratio,and the prediction score of short-term module are combined to judge whether the object predicted by the SiamDam algorithm is reliable.Secondly,it determines whether to start the re-detection module according to the reliability of the object,and if the predicted object is not reliable,the re-detection module based on co-occurrence statistics is started to relocate the object to obtain the accurate object position;Finally,the OTB100 was used to test the performance of the SiamCsLT algorithm.The experimental results show that the accuracy score of SiamCsLT algorithm is 0.902,the success rate score is 0.849,which is 6.0% and 6.1% higher than that of SiamDam algorithm,and 16.9% and 16.3% higher than that of SiamFC algorithm.The SiamCsLT algorithm has good tracking effect in long-term target tracking scenarios,and has a wide range of application prospects and practical value.