Research On Dangerous Situation Prediction Algorithm Of Intelligent Driving Based On Dynamic Traffic Scene Graph

A number of accidents caused by intelligent driving systems have drawn public attention to the safety of autonomous driving.In order to better understand the current driving scene and judge the safety degree of the current driving situation,this work will obtain more semantic interaction information in the driving scene to predict the dangerous situation in driving.The main research contents of this work are as follows:(1)Aiming at the lack of knowledge of the scene in the dangerous situation prediction,a scene graph model for the driving scene is proposed,which abstracts and interprets the relationship of each instance in the road,and extracts the spatial distribution and semantic interaction of objects,so that Unmanned vehicles can fully and effectively understand the scene.Aiming at the problems of lack of spatial information in the current scene graph generation task,insufficient learning of context information,and single evaluation index.The position coding method is used to embed the spatial information at a high level,and at the same time,the visual features of the object and the class label feature are combined to obtain a comprehensive object representation,and the improved multi-relation graph convolutional network is used to model the global context information.Based on the HDD data set,the performance of the scene graph generation algorithm was verified on the three subtasks of Pre Dcls,SGcls,and SGdet.By integrating multiple evaluation indicators in the current research and conducting comparative experiments,it is proved that the excessive reliance on a single indicator Recall@K in the previous research will make the evaluation results unable to describe the problem of model performance well.(2)In order to respond to the sudden danger scene on the verge of collision during driving,it is necessary to predict the dangerous situation.This work proposes an accident prediction model based on time-space relationship learning,which predicts the probability of traffic accidents through the video of the driving recorder.Most of the existing driving dangerous situation prediction methods only consider the characteristics of a single element on a frame,and do not consider the relationship between all traffic participants.For this reason,this work takes the scene graph in Chapter 3 as input,and models the semantic interaction information between targets in the form of a graph,and embeds the driving scene graph into the vector space and inputs it into the hazard prediction model to improve the model’s understanding of the scene.Based on the learning of the spatial relationship in the scene graph,the graph convolutional network is learned with the improved temporal relationship,and the potential spatio-temporal relationship features are learned and input into the Bayesian network to predict the occurrence of accidents.The model is trained on the newly released dangerous accident data set DoTA.The method can provide the probability of driving dangerous situation in real time,and predict the occurrence of driving danger with an accuracy rate of 89.14%and a recall rate of 80% before the accident occurs 1.8559 seconds.In order to verify the network generalization ability of the above-mentioned scenario graph-based dangerous situation prediction model,this work conducts a real vehicle test based on a smart car driving platform,introduces the hardware and software of the platform,and completes the performance test experiment in the campus driving scene.The experimental results verify the robustness of the dangerous situation prediction algorithm in this work.