Research Of Object Tracking Method Based On Local Sparse Appearance Model Under Complex Circumstance

Visual object tracking is one of the core component in computer vision,which can be widely applied to some practical applications such as self-driving cars,security and surveillance systems,intelligent transportation vision-based control.Although it has been heavily studied by many researchers for decades,and much progress has been made in recent years,robust visual tracking remains a challenging task because the appearance of object may be changed drastically while undergoing pose change,background clutter,illumination variation,occlusion and erratic motions.If the tracker cannot adapt to these changes,it will cause tracking failure.In recent years,sparse representation was used in the field of target tracking due to its good effect on handle the noise and model the appearance of object.There are three deficiencies in the existing methods of local sparse model:1)the contribution of different local patches when located the target object is not considered after dividing the target into several local patches,which cause the tracker lacks of discriminative power;2)the effectiveness of temporal context constraint do not considered when construct the object appearance model;3)the template update scheme which combines PCA and sparse representation will lose some information and the adaptive ability of tracker is poor.In order to solve the above three problems,this paper designs a weight-based local sparse appearance model,and then established a similarity of objects between frames as temporal context constraints.Furthermore,based on the improved template updating method,an adaptive target tracking algorithm based on local sparse representation is constructed.The main work of this article is as follows:(1)For the deficiencies of local sparse model using simple bounding box dividing mechanism to model the appearance of the target,this article presents an appearance model called weight-based local sparse appearance model,which first divided the target bounding box into several overlapped patches,and then calculate the patch reconstruction error as the weight to find relatively consistent and stable local patch in image.And finally,an alignment pooling operator taking the diagonal elements of the square matrix as final appearance feature.Tested on the 2015 PAMI benchmark data set show that after the weight operation of patches,the tracking algorithm can better restrain the influence of the occlusion and motion model.(2)Based on the foundation of the appearance model represented by weight-based local sparse representation,this article further construct a method that take the similarity of objects between frames as temporal context constraints.The interaction between the weight information and time information make the object representation more accurate.Moreover,currently template update based on subspace learning with sparse representation will lose some important discriminant information,we improved the method to make the sparsity only enforce to the noise(such as occlusion),so all of the principal component of the PCA matrix can be used to reconstruct the target and remove the impact of the noise portion.Finally,we tested the overall algorithm under public evaluation standard.Our method has great advantage in dealing with object tracking under complex scenes,such as partial occlusion,pose variation,motion blur and object size change.The proposed tracking algorithm performs favorably against the state-of-the-art methods on challenging sequences in terms of efficiency,accuracy and robustness.