Research On Robust Object Tracking Method Based On Superpixels And Feature Points

Object tracking is an important part of computer vision.It has been widely used in many fields,for example,intelligent surveillance,human-computer interaction,intelligent navigation and other fields.With the development of science and technology,the market demand for object tracking is increasing rapidly.Object tracking is still a challenging task because it is influenced by the deformation,occlusion and illumination of the object.With the development of the research,many part-based methods have been proposed.These methods have been proved to be robust to the influence of object deformation,occlusion and illumination.Superpixel is one of the most commonly used image regions,it is a mid-level visual cues.In object tracking,superpixel can preserve the inherent characteristics of the object,and it can provide more stable and reliable information for object tracking.This thesis studies the object tracking method based on superpixels and feature points,and the vehicle trajectory detection and analysis method based on this method.This thesis includes:(1)An object tracking method based on superpixels and feature points is proposed.This method adopts feature point matching and optical flow method to track the object candidate region.It adopts the superpixel method to over-segment the candidate region and constructs the superpixel feature description of the region.The superpixel feature description is composed of superpixel color histograms,feature points,and intrinsic structure relationships between feature points and superpixel centers.The final object region is obtained by superpixel matching and voting.The experimental results show that this method is robust to object deformation,occlusion and illumination.(2)Based on the above method,this thesis presents a vehicle trajectory detection and analysis method.This method uses the object detection and the similarity calculation to obtain the entry and departure state of the vehicle in a certain time window.The vehicle trajectory is obtained through the object tracking method proposed in this thesis.This method adopts an on-line self-learning method to establish a road trajectory information map,which is used for vehicle trajectory analysis.The experimental results show that this method can adaptively deal with complex and changeable road conditions.