Research And Implementation Of Pedestrian Detection And Location System Based On Binocular Image

In recent years,with the rapid development of computer vision,the demand for automatic detection and positioning of pedestrians to obtain the 3D coordinate information of pedestrians is increasing day by day,and accurate real-time binocular image stereo matching technology and image pedestrian detection technology have become the hotspot of academic research.In the presence of Gaussian noise in binocular images,Census transform of the traditional Semi-Global Matching algorithm is easy to produce the wrong value,resulting in reduced accuracy of stereo matching,in addition,the traditional Semi-Global Matching algorithm limit large changes of disparity in whole image,resulting in edge expansion and blur of disparity map.The traditional Fast-YOLO network is less capable of detecting pedestrians image with adjacent pedestrians.In order to solve the above problems,this paper focuses on Pedestrian Detection and Location System based on Binocular Image.Firstly,this paper presents a Semi-Global Matching algorithm based on SLIC superpixel segmentation.It mainly includes: 1)Using the method of mean filter preprocessing,the binocular image is pre-processed by mean filter to reduce the Gaussian noise and improve the accuracy of Census transform in the noisy environment;2)Using the method of SLIC superpixel segmentation,the image is done by SLIC superpixel segmentation,the smoothing coefficient of the energy function is dynamically set according to whether the adjacent pixels belong to the same superpixel,allows large changes of the disparity at the edge of the image texture.The experimental results show that the method of mean filter preprocessing effectively improves the accuracy of stereo matching in noise environment.The method of SLIC superpixel segmentation can effectively avoid the problem of edge expansion and blur of the disparity map and improves the accuracy of stereo matching.The algorithm has efficient calculation speed.Then,this paper presents a Pedestrian Detection Network based on Fast-YOLO,redesigning the network structure and network function: 1)Using the method of improving network resolution,the Pedestrian Detection Network has 14 × 14 network resolution;2)Using the method of single bounding box,the Pedestrian Detection Network only predicted one bounding box.The experimental results show that the Pedestrian DetectionNetwork based on Fast-YOLO greatly improves the detection ability of pedestrians image with adjacent pedestrians,improves the detection ability of pedestrian images with random pedestrians,and the algorithm has high detection speed.Finally,a Pedestrian Detection and Location System based on Binocular Image is designed and implemented in combination with the stereo matching algorithm and pedestrian detection algorithm.The system optimizes the algorithms in parallel and design the 3D display function.The experimental results show that the system can accurately output the 3D coordinate information of pedestrians,and can realize its 3D display.The system has high computational efficiency and can be used in real-time pedestrian detection and positioning products.