Research On Key Technologies Of Binocular Vision Detection And Ranging For Wheeled Autonomous Robot

Visual detection and visual ranging are the basic key technologies for wheeled robots to automatically perceive the object class and its distance in a scene,to independently synchronize positioning and mapping,and to independently plan and control walking paths.The accuracy of visual detection,the accuracy of distance and the timeliness of calculation directly determine the agility of the wheeled autonomous robot.This thesis aims to improve the accuracy of visual detection and visual ranging for wheeled autonomous robots,while taking into account the timeliness of calculation,from four aspects: the construction of high accuracy visual detection model,the calibration of binocular visual ranging system,binocular visual ranging,the implementation of visual detection and ranging system.The main research contents of this paper include:(1)The construction of YOLOv5(You Only Look Once v5)visual detection model incorporating characteristic tower attention mechanism and skip connection.Because YOLOv5 model is weak in capturing cross-scale features,ignoring the same scale similarity features and the gradient disappearance of the feature pyramid structure,it is easy to cause the model detection accuracy to be low.solve these problems,this paper replaces the feature pyramid structure in YOLOv5 with the feature pyramid attention mechanism model,adds skip connection structure at both ends of the feature pyramid model,and introduces Mish activation function.A YOLOv5 object detection network model with improved pyramid and skip connections is proposed.The comparison experiment results show that the detection accuracy,recall and F1 value of the visual detection model on the MS COCO dataset increase by 2.7,2.2 and 2.7 respectively,and the detection accuracy,recall and F1 value on the PASCAL VOC dataset increase by 6.1,1.6and 3.6 respectively.(2)Study on calibration of binocular vision ranging system.The principle of binocular ranging vision system is analyzed.The hardware model of binocular vision ranging system is selected and the software environment of the system is configured.A binocular vision ranging system is built.The binocular visual ranging system is calibrated by Zhang Zhengyou calibration method,and the parameters solved by the calibration experiment are used to correct the binocular visual ranging system,eliminating the camera lens distortion and completing the polar correction.(3)Improved adaptive aggregation stereo matching binocular vision ranging technology.The feature extraction module of AANet(Adaptive Aggregation Network)stereo matching network model has weak feature extraction ability in complex and weak texture areas,ignoring image feature consistency and effective field bias,which can easily lead to poor network prediction accuracy and high pixel error rate.The feature extraction module is improved by combining domain generalization method with feature optimization structure,and the feature similarity module is introduced to add content similarity constraint.An improved adaptive aggregated stereo matching network model is presented by incorporating multiscale variable attention into the cost aggregation module.The comparison results show that the pixel matching error rate of the stereo matching model is lower and the matching speed is faster.Based on this model,an improved adaptive aggregate stereo matching binocular visual ranging method is proposed.The experimental results show that the method can meet the requirements of visual ranging tasks.(4)Binocular vision detection and ranging system for wheeled autonomous robot.First,the hardware platform of the wheeled autonomous robot is built based on the raspberry pie center controller,the visual imaging device and the wheeled walking device.Then,the ROS(Robot Of System)is installed on the hardware platform,and the deployment of the above visual detection algorithm and visual distance measurement method is completed.In order to solve the problem of deployment difficulty due to the large parameters and computational consumption of deep learning algorithm,a combination of reducing the depth,width and pruning of the model network is used to achieve lightweight model.The experimental results show that the system can realize the visual detection and distance measuring functions of the wheeled autonomous robot,and the running speed of the system basically meets the requirements of real-time tasks.