Research Of Pedestrian Detection Based On Binocular Vision And CUDA

With the development of computer vision,detecting pedestrian through image acquisition and digital image processing has been widely used now.Because of the diversity of the pedestrian and the application scenes,accuracy and real-time performance of the existing pedestrian detection system is challenged.Aiming at narrow area's overlooking scenes,this thesis mainly studies on stereo matching algorithm,pedestrian detection and tracking algorithm,and construct real-time binocular pedestrian detection system combined with CUDA parallel computing technology,which have significant academic value and practical significance.At the overlooking scenes of pedestrian detection,the depth information is important for pedestrian recognition.For the binocular vision system,stereo matching is the most important step for getting the disparity map,whose precision is proportional to its algorithm complexity.Subject to the requirements of real-time,the conventional algorithm is difficult to obtain high accuracy.This thesis puts forward an improved algorithm based on the existing local stereo matching algorithm,which improves the accuracy of stereo matching with an adaptive weighted aggregation algorithm.In order to shorten the operation time of stereo matching,this thesis implements the algorithm with GPU computing platform and CUDA framework,and improves the adaptive weight algorithm with the introduction of coarse-to-fine idea.Experimental results show that it takes 25ms to process an image of 192*144 resolution,which can satisfy the demand of the common video processing.This thesis's improved algorithm ranks 102 in Middlebury tests.Compared with the original adaptive weight algorithm,this thesis's improved algorithm's operation time is only half of the original one while their accuracy is close.In order to improve the accuracy of the pedestrian detection,this thesis fully consider the useful information of the original image and the disparity image,and extract the movement feature,shape feature and depth feature as the basis of pedestrian recognition.With this target recognition algorithm,the accurate head position coordinates can be got and the missing or mistaken detection can hardly happen.In order to get multi-pedestrians' track,Kalman prediction algorithm has been added into target tracking link,which effectively improve the accuracy of target tracking.Finally,the image acquisition system and heterogeneous computing platform are constructed.In view of the stereo matching task computed with CUD A,this thesis identify the bottleneck through the performance test,and improve the GPU occupancy and data transmission efficiency through optimization.Considering the operation time of the major computing tasks and dependency relationship,this thesis puts forward a video processing improvement on heterogeneous computing platform.This improved processing method can complete the pedestrian detection task in a short time with full use of CPU and GPU computing resources.In the people-counting experiments conducted in indoor environment and industrial elevator,the accuracy reached 96%and 88%respectively.So,this system is accurate with good robustness.Finally,the research work of this thesis is summarized and the further research directions are given.