Research On Approach Assessment Of Road Traffic Danger Based On Deep Learning Model

In ADASs,the principal objective must be to uphold is that “the vehicles sent humans to the target without accident”,and one necessary technology for actualizing the goal is computer vision algorithms based on deep learning.However,there are few pieces of research on road traffic near-miss quantitative assessment,road traffic near-miss scenes are going unheeded,self-driving cars have little chance to learn how to respond to these situations.Thus it is essential to anticipate road traffic near-miss situation with computer vision algorithms.In the view of above problems,based on the traditional road traffic near-miss quantitative assessment studies,this dissertation provides real-time traffic near-miss quantitative evaluation information for dashcam with computer vision algorithms based on deep learning.The main work of the thesis are as follows:(1)Research on road user recognition method: Our research needs to identify the road usersinclude cars,buses,motor vehicles,non-motor vehicles and pedestrians from the drivingrecord video,that is,the 720×1280 picture in each frame.And the output of the method isthe corresponding road user convolution feature maps,which are the part of the input ofroad traffic near-miss feature portraits model and road traffic near-miss quantitativeassessment model.The research on recognition method using three different deep learningmodels: R-CNN,YOLO,and SSD.This dissertation analyzes the advantages anddisadvantages of three road user identification methods and lays a foundation for theestablishment of traffic near-miss feature portraits model and traffic near-miss quantitiveassessment model.(2)Research on road traffic near-miss feature portraits method: To better describe the road userswho may have accidents,this dissertation studies a generative model for road traffic near-miss feature portraits: DSA-based(Dynamic Spatial Attention)generative model andCAM-based(Class Activation Map)generative model.In the road traffic near-miss featureportraits method,we analyze and compare the advantages and disadvantages of the twomethods according to the road user identification result,and generate a set of weights forthe road users to describe the possible accident occurrence of different road users better.(3)Research on traffic near-miss quantitative assessment model: Based on the research resultsof road user identification method and road traffic near-miss feature portraits method,thisdissertation studies the road traffic near-miss quantitative assessment model based onseveral variants of the recurrent neural network,namely LSTM,QRNN,and QRNN-Inception based on Inception unit.This dissertation analyzes and compares the advantagesand disadvantages of the three models for more accurate and timely road traffic near-missquantitative assessment and provides early warning information.(4)Research on the framework of road traffic quantitative assessment: Combined with theresearch results of the above three research work,this dissertation studies the overallstructure of road traffic quantitative assessment.First,this dissertation discusses the lossfunction for early anticipation,including cross-entropy(CE),Adaptive Loss for EarlyAnticipation(AdaLEA)and Positive Favored Loss for Early Anticipation(PF-LEA).Secondly,a model optimization framework based on incentive neural network is establishedto synchronize the road traffic near-miss feature portraits model and the road traffic near-miss quantitative assessment model.Therefore,this dissertation integrates the road user recognition module based on YOLO,the traffic near-miss portraits generation module based on class activation map and the traffic nearmiss anticipation module(TNFP-GM)based on QRNN-Inception and PF-LEA,as well,it proposes a framework for optimizing TNPF-GM and QRNN-Inception simultaneously.