Robust Object Tracking Methods In Complex Environment

Visual object tracking is one of the research hotspots in the field of computer vision,especially in the background of the rapid development of artificial intelligence.As a key technology of connecting image processing and high-level semantic understanding,object tracking is widely used in human-computer interaction,intelligent transportation,video surveillance,video compression,automatic drive and many other fields.In the past few decades,researchers have proposed many tracking methods and achieved significant progress in this area.However,due to the complexity of the tracking environment and the variability of the target to be tracked,the object tracking technology is still a complex and open research topic,and there are still many problems to be solved for industrial-level applications.Long-time object tracking,especially in complex scenes is a very challenging task.Based on the classical object appearance model and the latest academic achievements in this field,this paper aims to propose effective and novel tracking methods to enhance the robustness of the object appearance model in complex environments,and thus to improve tracking accuracy and efficiency.The main contributions of this thesis are summarized as follows.Firstly,in order to achieve long-time robust object tracking in complex environment,this paper proposes to estimate target confidence based on super-pixel and bag-of-word model,and then combine foreground,background and depth information to estimate target state in the Bayesian framework.First,the super-pixel is used to segment the target foreground and background area.Then,the AP adaptive clustering method is employed to generate the dictionary of target foreground and background automatically.Lastly,the target confidence is estimated based on three kinds of features: local feature,global feature and depth feature.In the tracking process,the target model is updated with sparse representation to prevent the target model from degenerating and achieve long-time tracking.The experimental results show that the proposed method can effectively model the tracking target and realize the long-time and accurate tracking in complex scenes.Secondly,in order to achieve 3D collaborative tracking in multi-perspectives camera coordinates,this paper proposes to calibrate the 3D scene by combining multi-perspectives RGB cameras and depth cameras.Different from the monocular camera object tracking system,the proposed method comprehensively utilizes the appearance features of the target from different perspectives to construct a robust visual model.When the tracked target undergoes occlusion or sudden appearance change,the cameras from other perspectives still can estimate the target state.The proposed method provides a multi-perspectives object tracking framework that can be easily extended to more perspective tracking requirements.Experimental results show that the tracking method can effectively model the target appearance under multiple perspectives and complete the 3D target tracking across cameras.The proposed adaptive model updating strategy can effectively alleviate the model degradation and is suitable for long-time target tracking.Thirdly,for the current object tracking methods based on deep learning cannot effectively intergrade the background information to model target appearance,this paper effectively combines the background-aware correlation filter with deep learning and proposes an end-to-end deep background-aware object tracking method.First,the proposed method integrates the correlation filter as a special convolutional layer in the deep learning network,and then treats the target foreground and background as positive and negative samples to perform model training,which aims to enlarge the distance between the foreground response map and background response map.Experimental results show that the proposed method can effectively suppress the response of the target background and encourage the response of the target foreground,which prove that background information plays an important role in modeling the target appearance.The quantitative evaluation results on public available dataset demonstrate that the method can improve the performance of the original tracker significantly both in tracking accuracy and success rate.Lastly,in view of the problems of information redundancy and dimension disaster in deep learning features,some object tracking methods based on deep learning are not suitable for running on devices with limited computing power.Therefore,this paper uses an adaptive clustering method to analyze the intrinsic relationship between feature maps and select discriminative and representative features for object tracking.In order to utilize the description power from different layer features effectively,multiple trackers are trained using multiple layer features with different discriminative abilities to determine the target state jointly.Then,the weight terms of each tracker are updated dynamically based on decision errors.Experimental results show that the proposed method can reduce the feature dimension by about 80% without losing the tracking accuracy,and effectively improve the tracking efficiency.