Research On Moving Target Detection And Tracking Algorithm Based On Stereo Vision

With the continuous development of computer vision,moving object detection and tracking,which is one of the important branches of computer vision,has made great progress.On the other hand,along with the further development of social informatization,mass of cameras are equipped in every corner of society.In the meantime,a large number of video data producted by them still need to be further mined and utilized.In this context,pedestrian recognition,which is an active research area of moving object detection and tracking,has produced a stronger momentum of development.In this thesis,after analyzing the problem of pedestrian recognition algorithm based on monocular vision,the binocular stereo vision is introduced in moving object detection and tracking.Firstly,the principle of binocular stereo vision is described,and then the local stereo matching algorithm based on Census transform is studied.In stereo matching,the Census transform is a non-parameter-based similarity measurement function.While it has great advantages like good robustness,implementation simplicity,it also has some shortcomings,such as it is not sensitive to the image edge,and with the aggregate window becomes larger,its calculation speed slows down rapidly.Aiming at the existing problems of Census transform,this thesis presents a 2-bits Census transform based on the combination of pixel gray and gradient information,and a cost aggregation method based on the column window is adopted,which effectively improved the matching accuracy and matching speed.of the local stereo matching algorithm based on Census transform.The extraction of moving object region is the basic step of moving object detection,and background subtraction algorithm is the most commonly used method to extract moving object region,which also has the best performance.ViBe is a general background modeling algorithm,which is also one of the excellent background modeling algorithms proposed in recent years.By properties like single-frame-modeling,random sampling,it has obtained a great practical application performance.Considering the ViBe's disadvantage like sensitive to the shadow of moving object and bad performance on detecting long time static target,this thesis made some improvement and makes it more suitable for pedestrian detection.Pedestrian counting,as a common application scenario of moving pedestrian detection and tracking technology,has great practical value.The traditional pedestrian recognition is mainly relied on the pedestrian's movement property to realize the detection.However,when facing conditions like occlusion,overlapping and deformation,it has a pool performance.In this thesis,by using machine learning,a pedestrian classifier based on depth feature and support vector machine is introduced to accomplish the task of pedestrian recognition.Inspired by Histogram of Oriented Gradients,the pedestrian feature descriptor based on depth information is adopted.This feature is,then,used as the input of SVM for training.And it achieved a great recognition performance of pedestrians.At last of this thesis,a pedestrian counting system,including hardware system and software system,is constructed.The effectiveness of the proposed algorithm is verified by experiments.Finally,the research work which has been done in this thesis is summarized,and some aspects which can be further improved are proposed.