Research On Vehicle-End Perception Based On Monocular Object Detection And Depth Estimation

In recent years,based on the unprecedented development of computer computing hardware,breakthroughs in artificial intelligence research,and the promotion and encouragement of related policies,new opportunities have been brought for human-computer interaction and autonomous vehicles.Analogous to human driving a car requires "one look","two thinking" and "three actions".Autonomous driving technology is mainly composed of three major components,namely perception,planning and control.One of the indispensable sensors,which replaces human eyes,provides important image information for the self-driving car system,and obtains the position and distance information of traffic elements and target objects that appear around the car.Improving the accuracy and efficiency of vehicle object detection and depth estimation is a long-term challenge in the field of autonomous driving.This thesis studies the vehicle-side fusion perception system in the autonomous driving scenario,including object detection and depth estimation tasks.The main work is as follows:(1)Aiming at two categories of algorithm tasks,this paper collects domestic traffic scenes and produces 90000 image Datasets XCSY Datasets.The image annotation categories of the dataset include 7 categories: Vehicles,pedestrians,signs,road arrows,traffic lights,sidewalks,and faces contain 18 different types of elements.The labeling work is completed by the experimental team;(2)For the object detection task,the YOLOv5 object detection algorithm is improved to improve the accuracy of object detection at the vehicle end.The main improvements include:Backbone introduces CBAM attention mechanism;Neck introduces shallow feature layer reuse.The Head layer decouples the object detector and replaces the SIo U＿Loss of the bounding box loss function with the canonical vector Angle.The m AP performance of the proposed algorithm is improved by about 3%,reaching 72.48%;(3)For the depth estimation task,the Pack Net-sfm network algorithm is used to realize the prediction of the depth and distance information of the two-dimensional image,and the prediction accuracy of the model reaches 88%.The two types of algorithm experiment task reasoning achieve fast efficiency and high accuracy;(4)The output results of the two types of tasks are integrated into the image coordinate system,and the depth information of the traffic elements in the image frame is projected into the target frame of the object detection.In terms of decision-making,the pixel-level prediction error of the depth estimation algorithm is accepted within a certain range.The average depth distance information of all pixels in the object detection frame is used as the depth distance representative of the traffic element object,and finally the 2D position of the traffic element object in the input image frame and the depth distance from the object to the vehicle camera are output,which can better solve the problem of The identification and positioning of traffic elements and objects in autonomous driving road scenes are solved..