Research On Visual Object Tracking Algorithm In Complex Scenes

Visual object tracking is an important research direction in the field of computer vision.In object tracking,the initial state of the object is given in the first frame of the video.The position and scale information of the object is continuously estimated according to the limited prior information.With the continuous improvement of computing power and the development of big data technology,visual object tracking has been widely used in video surveillance,robotics,unmanned driving and other fields.In actual complex scenes,the changes of the target itself and the background environment such as target deformation,rotation,scale change,occlusion,background interference,illumination change,and the changes caused by shooting behaviors such as motion blur,out of view,and low resolution bring great challenges to the object tracking algorithm.How to maintain the accuracy and robustness of the object tracking algorithm in complex scenes is still a challenging problem.Making full use of the useful information generated in the tracking process,avoiding the interference of error information as far as possible,and enhancing the tolerance of the model to interference information are effective means to improve the robustness of the target tracking algorithm in complex scenes.Aiming at the problems of limited positive sample information,error accumulation and poor anti-interference performance of the model in complex tracking scenes,this dissertation uses ensemble learning and other methods to carry out the following researches in terms of sample quality,model correction and interference suppression.(1)Aiming at the problem that the correlation filter tracking algorithm is difficult to achieve long-term tracking in complex scenes such as target appearance change and frequent occlusion,a multi-model fusion long-term tracking framework is proposed,which combines a sample pollution mitigation mechanism and a real-time appearance model correction strategy.The framework alleviates the problem of sample contamination by re-detecting the target with an independent robust classifier.Aiming at the problem that the discrimination of the appearance model to the target decreases,the model with the strongest discrimination is selected from the multiple models updated online with different historical tracking results,and the contaminated appearance model is corrected.A criterion that can accurately measure the discrimination of the model to the target is proposed.Experimental results on video sequences with multiple challenging attributes such as occlusion and deformation demonstrate the effectiveness of the proposed tracking framework for most challenging scenes.Comprehensive experimental results on multiple video sequences demonstrate the advancement of the proposed tracking framework.(2)To solve the problem of poor adaptability of independent evaluation criteria and difficult coordination of joint evaluation criteria in multi-model fusion in complex scenes,a double fusion robust tracking framework is proposed.This algorithm uses three complementary evaluation criteria: maximum response value,PSR and energy function.When the final tracking results given by the three evaluation criteria are different in complex scenes,the framework uses motion dynamic model and forward-backward analysis for re-judgment.The motion dynamics and forward-backward consistency assumptions are introduced to complement the appearance information to obtain reliable tracking results.The experimental results on multiple datasets such as OTB and TC-128 prove the robustness of the proposed algorithm in complex scenes,and the comparison experimental results with advanced algorithms show the effectiveness and superiority of the proposed algorithm.(3)Aiming at the problem that the existing tracking algorithms have poor adaptability to target occlusion and lack of effective occlusion detection strategies,a general spatio-temporal occlusion perception strategy is proposed.The proposed occlusion perception strategy detects the occlusion from the perspective of time domain and spatial domain respectively,and combines the two to give the final occlusion detection result.The detected occlusion results are used to implement a conservative online model update strategy,which avoids the contamination of the apparent model by occluded samples,and also alleviates the problem of model overfitting commonly encountered in tracking algorithms.Extensive experiments on several datasets including OTB-2015,OTB-2013,TC-128,UAV123 and UAV20 L show that the proposed strategy can greatly improve the performance of tracking algorithms.The robust performance of the proposed strategy in the occlusion scene indicates that the specialized analysis and processing of the occlusion problem are effective.(4)In order to solve the problem that the target tracking algorithm based on Siamese network depends on the first frame target template in the tracking process and cannot adapt to the change of target appearance in complex scenes,this dissertation proposes a multi-template tracking algorithm based on Siamese network with interference awareness and reliability awareness.Inspired by the correlation filter tracking algorithm,the template of the tracking algorithm based on Siamese network is updated online.At the same time,in order to avoid error information when the estimation information of subsequent frames is introduced,and reduce the discrimination of the template to the target,the strategies of reliable sample perception,distractor perception and multi-template fusion are proposed.The proposed tracking algorithm not only improves the robustness in complex scenes,but also maintains the discrimination of the template to the target.The experimental results based on multiple data sets show the effectiveness and advancement of the proposed algorithm.Aiming at improving the robustness of target tracking algorithm in complex tracking scenarios,this dissertation proposes a series of innovative methods from the perspective of reducing error samples and preventing pollution models,which effectively improves the performance of target tracking algorithm and has certain guiding significance for the application of target tracking algorithm in actual scenarios.