Real-time Monocular Distance Measurement Research In Automated Driving Scene

As one of the most important inventions since the 18th century,automobile has greatly improved the pace of people's lives and promoted the development process of civilization.However,when people enjoy the convenience of cars,also suffer from traffic accidents.According to incomplete statistics,no less than 10 million people are killed or injured in traffic accidents worldwide every year.In order to prevent traffic accidents,current vehicles must maintain a safe distance from pedestrian vehicles in front.As an important part of the self-driving system,it is also essential to improve the reliability and safety of the self-driving system.Target detection is the basis of target ranging,and this paper proposes a pedestrian detection algorithm based on lightweight YOLOv3 for the autonomous scene.Combined with pedestrian test results,an similar triangle distance measurement algorithm with pitch?pitch angle?and yaw?yaw angle?is proposed.With the in-depth study of the deep estimation task,a single-view depth estimation and detection algorithm?Monocular Depth Detection Network,MDDN?based on deep-aware collaborative network is proposed.Aiming at the problem that the current target detection algorithm is not up to real-time and the accuracy is low,this paper combines with the scene of pedestrian detection,basing on the common target detection algorithm YOLOv3,a lightweight,high-precision pedestrian detection algorithm is proposed;the limitations of the existing network input dimension design are discussed,and an effective non-deformation network input strategy is proposed;The model compression and the design of lightweight network are used to reduce the calculation and parameter amount of the model;the feature pyramid module is combined to further enhance the characteristic expression ability of the network;the anchor distribution mechanism of the network output layer is changed to eliminate the long tail distribution deviation of the data set label,and effectively improve the detection accuracy of targets.The final lightweight network on the BDD 100K validation set,pedestrian detection of the AP₅₀?Average Precision?reached 70.32%,mAP?Mean Average Precision?51.5%,The mAP is 5.9%higher than that of YOLOv3 algorithm,which has better detection robustness,and the model parameters and calculation are 5.32M and 7.12GFlops respectively.The time to measure a picture is about 10 milliseconds.Pedestrian targets in the automatic driving scene are diverse in form and size,and it is not possible to use regression modeling methods to fit a common model and measure all pedestrians,and the widest distance measurement model currently used is a geometric derivation method based on the camera imaging model and geometric perspective relationship.Therefore,this paper makes a detailed analysis of the commonly used similar triangle distance measurement algorithm,and proposes an improved similar triangle distance measurement algorithm with pitch and yaw,and on the basis of pedestrian detection,the pedestrian distance measurement system lightweight detection model is realized,with an average error of less than 6%within 90 meters,The frame rate on the FPGA is about 20,which satisfies the real-time pedestrian distance measurement.The traditional visual distance measuring algorithm is used to detect and then measure distance,and the measuring results are greatly affected by the test results.Therefore,this paper innovatively combines the target detection task with the depth estimation task,presents an end-to-end multi-task model of real-time monocular depth estimation and detection,and carries on pedestrian vehicle detection and depth estimation for a given single picture,and finally achieves the distance measurement.Reference to the structure of HRNet model,a lightweight multi-layer Internet backbone is designed to preserve high-resolution feature maps;The sharing strategy is automatically learned by dynamic routing through the horizontal sharing unit,and the target detection sub-network and the depth estimation sub-network are designed respectively.The algorithm realizes the distance measurement of pedestrian vehicles based on depth estimation,and on the KITTI data set,the accuracy of vehicle detection and pedestrian detection is 81.41%AP and 60.07%AP respectively,and the depth estimation index achieved slightly higher accuracy than the latest algorithm,with a relative error of 13.4%within 80 meters,the average error of the distance measurement system within 50 meters is 12.54%.