Research On Target Recognition And Location Based On Stereo Vision

Target recognition and positioning based on stereo vision is a hot issue in areas such as intelligent transportation and intelligent driving.The binocular camera recognition and positioning system installed on the UAV can obtain the status and relative position information of vehicles and other traffic targets early,and real-time monitoring and positioning of vehicles and other targets on the road from a high-altitude perspective,which is convenient for obtaining large-scale natural driving primitives Data and its driving model.The research is based on the recognition and positioning technology of vehicles and other targets under the unmanned aerial vehicle moving platform.The research focus of this article is set as: image enhancement and binarization of the collected video frame images;the introduction of traditional algorithms and volume The technical scheme combined with the product neural network is used to identify the target;the binocular positioning technology is used to analyze the three-dimensional position of the target,and the proposed algorithm scheme is realized by running on the embedded device,The specific research content is as follows:(1)In terms of target acquisition,first collect video data with binocular cameras,decompose the video sequence into picture data,and then perform median filtering on the image to reduce noise and eliminate image noise.Then the image is enhanced by the antilog function enhancement method.Finally,the traditional Mean shift algorithm and the OTSU algorithm are used to conduct experiments respectively,and it is determined that the OTSU algorithm is used to binarize the image and save the calculation space.(2)Use the convolutional neural network to identify the target: change the conventional background feature extraction network in the original SSD algorithm.This paper introduces the latest MobileNetV3 lightweight network to replace VGG16 as the feature extraction part,and the SSD algorithm is built The MobileNetV3-SSD network model has been compared and analyzed in many aspects with the current better detection and recognition algorithms.The experimental results show that the Loss function declines rapidly.MAP has a slight advantage over the one-stage detection algorithm,and loses some accuracy compared to the two-stage detection algorithm.However,it has reached 32 FPS in terms of detection speed,which has obvious advantages,and it is also verified.The effectiveness of the proposed network model algorithm is discussed.(3)In the research of three-dimensional positioning,a semi-global algorithm is used for distance measurement and positioning.Aiming at the mismatch and poor real-time performance of the semi-global stereo matching SGBM algorithm,this paper uses similar organizational constraints to improve the RANSAC algorithm,speeds up the calculation speed of the RANSAC algorithm,uses the Fast algorithm to extract feature points in the image and uses Euclidean Distance performs feature matching.Transplant and adapt the MobileNetV3-SSD network model trained on the server to the NVIDIA Jetson AGX Xavier embedded device,so that the entire algorithm system can run smoothly on the UAV embedded platform device,experiments were conducted on the completed UAV platform.The final depth estimation shows that the error value is below 6%,which verifies the effectiveness and practicability of the algorithm.