Research On Camera-in-the-Loop Test Method For Pedestrian Collision Avoidance System Of Intelligent Vehicle

Intelligent vehicle pedestrian avoidance system is an advanced driving assistance system(ADAS)based on visual information designed to avoid or reduce vehicle injuries to pedestrians.Since road test is difficult to reproduce complex working conditions,camera-in-the-loop test can make up for this deficiency by embedding real camera hardware and constructing virtual scenarios.Therefore,the establishment of a camera-in-the-loop test platform to conduct in-depth research on the influencing factors of camera imaging,and build virtual test scenarios based on this,explore accelerated test methods,and conduct automated tests on intelligent vehicle pedestrian collision avoidance system based on visual information is the important research content in the test field of intelligent vehicles in the future,and is also the prerequisite and foundation for the realization of autonomous vehicles.Supported on the National Key R&D Program of China(2018YFB0105103),this paper has carried out the research on the camera-in-the-loop test method.A camera-inthe-loop test platform has been established,which is verified the effectiveness and high confidence of the platform through real-vehicle data collection and product-level intelligent vehicles.A pedestrian collision avoidance system is designed based on convolutional neural networks.The virtual test scenarios are constructed by analyzing the influences of camera imaging.Based on the combinatorial test theory,a method of generating accelerated test scenarios is designed.Finally,automated test on the test platform and experimental results analysis are carried out.(1)Construction of camera-in-the-loop test platformThe camera-in-the-loop test platform is designed,and based on the virtual scenario projection method,the camera-in-the-loop test platform is built by selecting the monitor,vehicle-mounted camera,video box,Prescan,Matlab/Simulink and other software and hardware.After calibration and calculation,the Mobileye camera is fixed on the test platform to verify the functionality and scenario simulation by black box test.And the confidence of the test platform is validated through the comparison test between actual vehicle and platform.In order to improve the test efficiency,an automated test process based on this test platform is proposed.(2)Design of pedestrian collision avoidance system for intelligent vehicleThe architecture of the intelligent vehicle pedestrian collision avoidance system is proposed,and the monocular vision intelligent vehicle pedestrian collision avoidance system based on Yolo-v3 and Yolo-v4 detection networks is constructed respectively.The public data set COCO and Caltech are selected to establish a joint training set,and the two detection networks are trained respectively,and the longitudinal distance between people and vehicles is calculated according to the principle of monocular distance measurement.The active collision avoidance module based on the minimum safe distance model is designed,and the communication between two modules is completed through the UDP protocol.The system can complete the pedestrian collision avoidance function on the test platform.(3)Research on method of generating test scenarios for pedestrian collision avoidance systemAccording to the camera-in-the-loop test function requirements,the influence of light,rain,snow,fog and motion on camera imaging is analyzed through the camera module,and the test scenario parameter design is carried out.The speed of the vehicle,the light,the weather category,the color of pedestrians and the shadows of roadside trees are used as scenario parameters,and discrete values are selected according to the test requirements,and the specific test scenario design is completed.Since the number of traversal test cases will increase exponentially with the increase of the number of scenario parameters and specific values,in order to solve such problems,this paper proposes a method of generating combinatorial test cases based on the greedy algorithm.(4)Test experiments and research on evaluation methodsBased on the camera-in-the-loop test platform,the pedestrian collision avoidance system,target detection algorithm and the generating method of test scenarios are automatically tested and analyzed.The chi-square and two-factor variance method were used to analyze the test results of the pedestrian collision avoidance system.For the target detection algorithm,the dynamic performance index of the algorithm-the detection accuracy of the first detection distance and the minimum safe distance is proposed,and the common target detection index m AP and FPS are combined for joint evaluation.For the scenario generation method,compared with the traversal test,the combinatorial test based on the greedy algorithm improves the test speed by nearly 12times;at the same time,the frequency of each scenario parameter value is roughly the same as the traverse test,and the experimental results based on the combinatorial test are carried out by a two-factor method Analysis,the analysis conclusion is basically the same as the traversal test;the experimental results FPS and the first detection distance obtained by the two test case generation methods are normal,and the distribution conditions are almost completely consistent.