Research On Background Information Suppression And Utilization In Object Tracking

Visual object tracking,which plays a significant role in computer vision,is widely applied in surveillance,navigation,the military,aerospace,and so on.In computer vision,many image information processing methods have achieved great development.Therefore,researches focus on video information processing.Object tracking can represent structural data,and analyze the development of video information.Object tracking can promote the progress of computer vision.Currently,object tracking focus on general tracking method for unknown target in unsupervised scenes.Object tracking involves predicting variation in the position of a target in image sequences with a given initial state.Tracking environments are complex and vary with time.For actual application,object tracking requires robustness,adaptivity,and real-time performance.The main challenges in object tracking includes illumination variation,scale variation,occlusion,deformation,motion blur,fast motion,in-plane rotation,out-of-plane rotation,out-of-view,background clutter,and low resolution.Visual tracking consists of four parts: motion model,feature extraction,observation model,and tracking prediction.Motion model predicts target motion and provides candidates.Feature extraction utilizes vectored data to describe the target and the candidates.Observation model evaluates the similarity between the candidates and the target.Tracking prediction locates the target with observed results.For object tracking,background information has two functions.The one is to disturb the locating of the target.Suppressing background can reduce the mistake which treats the background as the target.The other is to help locate the target.The background information surrounding around the target can be used to infer the position of the target.This paper has researched the suppression and utilization of background information in visual tracking from perspectives of motion model,feature extraction,observation model,and tracking prediction.A complete tracking system involving background information is established.The main innovations are as follows.(1)The long-and short-term motion model with collaboratively modeling the target and its background is proposed.The target and background are separated and tracked individually.Their tracking results are modeled collaboratively,and the change rule of the relationship between the target and the background is learned.The short-term motion model predicts the motion of the target in a large region,and evaluates the stability of the object tracking.When the tracking stability is low,long-term motion model generates several candidate regions in the whole image to find the target.Short-term motion model enables common trackers track the target in large regions,and can solve fast motion.Long-term motion model re-detects the target after loss,and recovers the tracking process.Experimental results show that proposed motion model can deal with target loss and achieve promising performance in long-time videos.(2)The target and background feature extraction method for visual tracking by structurally optimizing pre-trained CNN is proposed.The pre-trained CNN is optimized to keep features suitable for tracking problem,as well as simplicity.The compressed CNN can suppress background and generate better features for tracking.The proposed method consists of the PCACS(Principal Component Analysis with Channel Selection)dimensionality reduce method,the channel selection method involving solving the TVBO(Target Variation and Background Output)objective function,and the single sample weights reconstruction method.The PCACS takes the information amount and tracking error of feature maps into consideration to obtain good low-dimensionality features from the last convolutional layer for tracking.Minimizing the TVBO can select representative channels,keep the target information,suppress background information,and optimize the network structure.The single sample weights reconstruction method can learn weights for remaining filters to reduce the loss of target information.Experimental results show that the features from compressed network are stable and have strong ability to distinguish the target from the background.Furthermore,the tracking precision and efficiency are increased,and the computational requirements are reduced.(3)The multi-level context-adaptive correlation filters observation model is proposed.The discriminative correlation filter(DCF)has shown impressive performance and high speed in visual tracking.Context has two functions in DCF: addressing the disturbance in target locating,and supplying cues for locating the target within the context.Firstly,the multi-level context-adaptive tracking(MCAT)approach introduces a multi-level context representation called a context pyramid to exploit the relationship between the target and its context for better visual tracking.Secondly,for each level of the context pyramid,the effect of context in DCF learning and tracking is controlled using context-adaptive spatial windows.An accurate target model can thereby be learned,even when the background clutter is severe.Moreover,the target can be more easily tracked when the background is weakened by the spatial window.Thirdly,a robust prediction of the target position is obtained with the multi-level structure of the context pyramid.Experimental results showed that,with conventional hand-crafted features,our tracker provides state-of-the-art performance on OTB100 comparable to those of deep-learning-based trackers and keeps the tracking efficiency.(4)The tracking prediction method with auxiliary objects in backgrounds is proposed.The dynamic background of images contains effective information that is favorable for object tracking.This paper uses auxiliary objects describing the target background.Motion dependencies between auxiliary objects and target itself are established first.Then the tracking results of auxiliary objects predict the position of the target.The position results predicted by auxiliary objects and by the basic tracker are fused to get a better result.Meanwhile,parameters updating is done according to the tracking result.Extensive experimental results show that auxiliary objects are effective for tracking the target.Comparison with other trackers proves the method's better robustness and precision.At last,the characteristics of background suppression and utilization in different tracking parts are discussed.This paper researches the functions of background information in motion model,feature extraction,observation model,and tracking prediction for object tracking.The proposed methods compose a complete tracking framework,and enhance the robustness,the adaptivity,and the real-time performance in object tracking.