Research On Siamese Network Tracking Algorithm Based On Target-distractor Aware

Object tracking is widely used in video monitoring,automatic driving and vehicle automatic navigation scenarios,and thus become the basic research and research in the field of computer vision research hot spots.At the same time,it is a challenging research task because of the arbitrariness of the target and the complexity of the video scene in the target tracking task.In recent years,the strong expressability of deep features provides strong support for the further development of object tracking,but the early tracking methods based on deep learning are difficult to get rid of the slow tracking speed caused by the deep features of large data.The tracking method based on Siamese network can not only benefit from deep learning and obtain good accuracy,but also has good speed,so it has become a very hot research branch in the field of object tracking.However,the object tracking method based on Siamese network is often difficult to recognize in the face of similar distractor,one of the important reasons is that the tracking method based on Siamese network only uses the given target template to measure the similarity,ignoring the further mining of target information.Another important reason is that Siamese-based tracking methods take little or no account of the discriminant information provided by the use of background.This paper mainly studies how to introduce background information into the Siamese tracking framework to prevent tracking failure in the face of similar distractor.The main research work and innovations are described as follows:(1)A target-distractor aware object tracking with adaptive appearance model fusion was proposed.Firstly,a distractor model is designed to make full use of the valuable cues provided by the background.It determines the weight of each distractor by the similarity between distractor and target,making the model focus on the distractor with high similarity to target.Secondly,the distractor model transformation model is constructed,the influence of the distractor model on the appearance modeling is ranked,and the similarity relationship between the background distractors and the target is mined by regularized linear regression,so as to effectively control the influence of the distractor model.Finally,in order to improve the discriminant ability of the appearance model,the model was unified into an adaptive distraught appearance suppression model.According to the intensity of the distractor,the distractor model was selectively used to suppress the distraught.The estimated target position is successfully prevented from drifting to the interference by mistake.(2)A target-distractor aware object tracking with discriminative enhancement learning loss is proposed to learn target representation,which can better identify targets from complex scenes.Firstly,to widen the gap between foreground and background,a discriminative enhancement learning loss is designed.The sensitivity of each channel feature to the target representation can be evaluated more accurately by highlighting the difficult negative samples similar to the target and shrinking the easy negative samples of the pure background.In order to obtain more robust target representation,a target-distractor aware strategy is proposed.By activating target-sensitivity and distractor-silence feature,target specific feature space can be obtained,thus effectively preventing tracking robustness from being affected by distractor-sensitivity and distractor-silence feature.Finally,the object representation of target-distractor aware is combined with Siamese matching network to achieve robust and real-time visual tracking.In order to verify the effectiveness of the proposed method,experiments are carried out on several object tracking datasets OTB-2013,OTB-2015,TC-128,UAV-123,VOT2016 and Lasot,and the experimental results show that the proposed two methods not only achieve good results in the case of background clutter,but also achieve good performance.Tracking performance also improved in other challenging scenarios.