Research On Object Tracking Based On Superpixel And Sparse Representation

Object tracking is a hot research topic in the field of computer vision.It has been widely applied in many scenes,such as intelligent transportation,smart military,intelligent medical treatment,intelligent video surveillance etc.Object tracking technology of further accuracy and robustness,could be able to provide better support for the follow-up image understanding.In addition,intelligent object positioning and tracking can effectively increase the existed image system efficiency,reduce human errors and losses,thus reduce the system operating costs.Therefore,object tracking has certain theoretical significances and application values.Existing target tracking methods are facing some challenges.such as partial occlusion,illumination or deformation,movement,and so on.By using of the sparse representation of visual features and super-pixel partition to maintain the edge of the target,an object tracking method based on super-pixel segmentation and sparse representation is proposed.The main work and achievements are summarized as follows:(1)When the target appearance change is caused by occlusion,illumination and movement,the problems of low tracking accuracy and poor robustness have often occurred.Aiming at solving these problems,an object tracking method based on super pixel and sparse representation is proposed.Firstly,super pixel was adopted to segment an image into blocks,image features were extracted and the result was structured sparse representation,and model was built;Then,the sparse coefficient and residual information were calculated and ranked,which were respectively used for preliminary target estimation and test;Subsequently,sparse coefficient and residual information were fused by using of a similarity function;Finally,the target template was updated,and the impact of change on the tracking target representation was reduced.Experimental results show that the average central error of our method is 7.6 pixels.In robust experiments,the tracking effect and accuracy of this method under noise attcack are still higher.Therefore,our method has good effect,and is robust too.(2)To overcome the influence of the number of super-pixels segmentation on object tracking accuracy and robustness,an object tracking method based on two-level super-pixel segmentation is proposed.Firstly,bilateral filter was employed to denoise.Secondly,the image was processed by coarse-grained super-pixel segmentation method,and the model based on the coarse super-pixels was built.Then,based on the results of coarse-grained super-pixel segmentation,fine-grained super-pixel segmentation was performed to calculate the confidence map.Finally,the representation model and the confidence map were updated.Experimental results show that the average central error of our method is only 7.5 pixels,and the error rate is still lower under noise attack,.Our method improves object tracking accuracy and robustness.(3)To overcome the influences of drift and occlusion,which traditional update methods cannot solve properly.,the feedback-based update strategy was proposed Firstly,we get the initial state and feedback state.Secondly,the Jaccard distance between the initial state and the feedback state was calculated and the fitness of the model was evaluated.Finally,the threshold was set,compare the fitness of the model with the threshold,and according to different situations,the template and confidence map was updated.Experimental results show that the updating strategy of this method improves the effect of template updating and improves the tracking accuracy.