Research On Potential Risk Assessment Method Of View Obstruction Scenes Driven By Data And Knowledge

Autonomous driving technology has made breakthroughs in the last decade due to the application of deep learning of artificial intelligence,but there are still many questions about whether driverless cars can drive safely in complex mixed traffic environments in cities where they share the road with humans.Especially,the problem of potential risks caused by the obstruction of self-vehicle sensors.How to accurately assess the potential risks contained in various types of view-obscuring scenarios themselves has still not been effectively addressed.To address this problem,this thesis proposes a data and knowledge driven method for assessing the potential risk of view-obscuring scenes based on the cognitive decision-making process of human drivers,which includes textual caption of multidimensional scene information considering fine-grained attributes and risk assessment based on inductive and analogical reasoning cognition.The main research work is as follows:The construction of a multidimensional scene image description model considering fine-grained attributes.To address the problem that the traditional video/image text description method contains single scene information and lacks the target object motion state,this thesis proposes a DSLT-C(Dynamic+Scene+Location+Time Caption)model,multi-dimensional semantic text information dynamic,image,macro location and time text information.Firstly,a detailed basic caption of the traffic scene is achieved by coding and decoding image caption networks Faster RCNN and LSTM,the targets in the image are described in text;after that,a bilinear recognition network for fine-grained recognition of vehicle models is trained to classify vehicles,and the recognition results are fused into the description text by a rule-based method;a combination of YOLO v5 target detection network,DEEP SORT tracking algorithm and monocular vision ranging principle to calculate the motion state of the occluded object and its relative position,and map its motion state and position results into 9 regions of the autonomous vehicle;use YOLO v5 to detect traffic signs and traffic lights;meanwhile,obtain the macro location text information of the scene by calling the GPS interface and output it as text;finally,through the IOU+anchor frame algorithm to achieve the uniformity of image description and motion state.The model solves the problem that the traditional video/image text description method contains single scene information,and can provide more complete and detailed environment perception information for the autonomous driving system.Traffic scene risk assessment based on inductive reasoning and analogical reasoning cognition.To address the problems of poor migration,overly complex model and poor timeliness of traditional risk assessment methods for view-obscured areas,this thesis proposes a data-plus-knowledge-driven risk assessment model for view-obscured scenes,which conducts risk assessment from the perspective of abstract text semantic information.For the general field of view obstruction scenario,a rule-based mapping method of scene semantic information is given to simulate the inductive reasoning mechanism of human drivers,which maps the complex and diverse scene text information into a fixed single feature text,and completes the risk assessment of the current scene based on the knowledge base of potential hazard scene risk determination established in the early stage.For the long-tailed vision-obscuring scenes,the analogical reasoning mechanism of human drivers is simulated,and the risk assessment of the current scenes is completed by solving the semantic similarity between the encountered scene caption text and other scene caption texts in the knowledge base.The risk scenario migration cognitive experiments show that the proposed method is migratory and applicable to a wider range of scenarios;the scenario risk assessment and comparison experiments show that the accuracy of the proposed method outperforms other methods in the field of potential risk assessment.The final analogical reasoning scenario risk assessment experiments illustrate the soundness of this paper’s approach in risk assessment.End-to-end real vehicle experiments.Traffic scene image data are collected in situ using a family car installed with a monocular camera,and a sample dataset is extracted to verify the effectiveness and real-time performance of the proposed method in this thesis.Real-time end-to-end output is achieved from acquiring scene text information to cognitive risk assessment based on inductive reasoning and analogical reasoning.Through the evaluation of common scene types and comparison with human drivers’ judgment of scene risks,the proposed method is proved to be effective for potential risk assessment of in-vehicle sensor field of view occlusion scenes.