Research On Intelligent Vehicles 3D Object Detection Based On Binocular Vision

Intelligent Vehicles perceive the state of the vehicle itself,roads,pedestrians,traffic signals,traffic signs,traffic conditions,surrounding vehicles,etc.through its perception system.Implementation of these functions requires lidar,millimeter-wave radar,cameras,and ultrasonic sensors to get surrounding information.Lidar is expensive and lacks semantic information.Monocular cameras lack depth information,and binocular vision can recover the three-dimensional information of the target through geometric relationships.Therefore,the study of accurate and fast binocular vision three-dimensional target detection algorithms is of great significance to the recognition and positioning of targets in complex environments.This paper takes the vehicle recognition in the road traffic environment as the research object,aiming at the problems of low target recognition accuracy,poor real-time detection of binocular vision three-dimensional target,and large amount of calculation.Based on Faster-RCNN,combined with deep separable convolution,a three-dimensional target detection network is established to recognition and locate vehicles quickly and accurately.This article mainly does the following research work:(1)Build a binocular vision system.Understand the theory of camera imaging and binocular distance measurement,and build an algorithm development platform and binocular vision system from the hardware and software levels.The existing calibration algorithm is used to complete the calibration of the camera,and the horizontal correction and distortion correction of the image are realized through the calibration results.(2)Establish a three-dimensional target detection network Ds-RCNN based on deep separable convolution and Faster-RCNN.A multi-scale feature extraction network based on deep separable convolution is used to solve the problem of large calculation amount and poor real-time performance,and improve the recognition of small targets;improve RPN to binocular regional regression,extract regions of interest to improve algorithm accuracy;Use the method of key point regression to get the three-dimensional information of the vehicle;propose an optimization method based on luminosity consistency,and optimize the results in a coarse-to-fine way.(3)Use the public data set to detect the performance of the algorithm and compare it with different algorithms.Test the rationality of the algorithm design by changing the algorithm structure.Deploy the binocular vision system to the intelligent networked car,and test the local environment and design experiments to verify the effectiveness and robustness of the algorithm.