Dangerous Scene Test Platform Construction Based On Natural Driving Data

At present,both China and abroad are vigorously developing Advanced Driving Assistance System(ADAS),which aims to assist drivers,avoid human error and improve Driving safety.To judge whether ADAS system achieves its development purpose,it needs to be tested on the test platform.This paper mainly builds the test platform of Adaptive Cruise Control(ACC)in the advanced driving assistance system in the dangerous scene and the platform is built with the help of simulation software PreScan and MATLAB.The dangerous scene is extracted and classified from a large number of natural driving data,so the extraction and classification of dangerous scene is the basis for building the test platform.The main research work and achievements are as follows:(1)Data acquisition and pretreatment.Collected data is analyzed into intuitively readable data according to different decoding methods and saved into Excel.Clean the parsed data by deleting invalid data,adding empty data by cubic spline interpolation and filtering noise,so as to facilitate subsequent use.(2)Extraction of dangerous scenes.The processed data were analyzed to determine the danger features of the scene,including longitudinal acceleration,THW and TTC.The characteristic parameter threshold is determined by referring to many foreign literatures and combining with the characteristics of its own data.According to the selected feature parameter screening criteria,the dangerous scenes were extracted,but only 61.97% of the screened fragments were dangerous.Among them,the scene was mistakenly screened for traffic light problems,road problems and driver driving problems.Different judgment conditions were added for different problems,and the screening model was improved.Accuracy of the improved screening model reached 98.92%,and PPV was increased to 81.72%.(3)Classification of dangerous scenes.The extracted dangerous scenarios were classified.Decision tree,random forest and LightGBM were used for scene classification.The first one and second one methods adopted grid search,and the third one adopted bayesian optimization to select the optimal parameters and compared the classification results under the optimal super parameters.The classification results show that the decision tree is easy to overfit.The accuracy of random forest classification is not ideal.LightGBM does not overfit,and the classification accuracy is the highest among the three methods.Therefore,LightGBM is selected as the subsequent scene classification method.(4)Dangerous scene test.Firstly,the driving environment of dangerous scenes were built in PreScan.Then the initial position and running state of the obstacles,as well as the initial position and state of the vehicle were introduced to complete the test platform construction of the scene.The scene can be observed the moving state,surrounding environment of the vehicle and obstacles from varities of dimensions,like the perspective of the driver,the top and the Human View.To build the test platform of joint ACC model in MATLAB/Simulink to control the vehicle movement,and compare the difference of vehicle motion with driver control and system control to achieve the purpose of testing the ACC system under the dangerous scene.The platform mainly tested three dangerous scenarios: the leading car cut in,the leading car brake and the braking of the target vehicle when the subject car cut into the leading lane.The test results show that the dangerous scene test platform can well restore the real driving environment,and can realize the ADAS test under the actual unrecoverable dangerous driving situation.