Research On Road Traffic Environment Perception And Driving Risk Assessment Method

With the sustained development of China's social economy,usage of automobiles show a rapid growth trend,which results in the increase in the number of traffic accidents and the number of casualties.In recent years,with the progress of science,a large number of advanced electronic aids have been applied to automobiles,e.g.,electronic stability control and advanced braking system,pedestrian protection,lane deviation and blind zone hazard warning functions.However,due to the development of assistant driving technology,many technical details have not yet been fully resolved.At present,the vehicle electronic assistant system cannot fully realize intelligent assistant driving.In such condition,it is impossible to minimize human-induced traffic accidents by early warning,warning and taking over vehicle control according to different driving risk levels.In order to achieve the goal of reducing traffic accidents or their severity through advanced auxiliary driving systems,the primary premise is to realize the perception of road traffic environment and real-time traffic risk assessment based on multi-sensor technology.Therefore,on the basis of previous research results,this paper carries out research on road scene classification,road scene segmentation,lane marking recognition and vehicle driving risk assessment methods.This paper has achieved the following research results.1)A road scene classification method based on image features is proposedThe road scene images are divided into three categories: urban road,ordinary highway and freeway.The differences of the features in the three kinds of scene images are analyzed.Firstly,the scene image is preprocessed and the feature extraction of guide sign,intersection sign pillar and entropy in the scene image is carried out.The VGG-Net model is fine-tuned according to scene image features to obtain scene recognition prediction values.At the same time,the posterior probability of features is calculated according to the prior probability values of different features in different scenes.Accurate recognition of road scenes is carried out by combining scene prediction value with Bayesian classification method.The method is validated by using domestic road scene data sets,and the recognition accuracy can reach93%.2)A road scene segmentation method based on multi-source data fusion is proposed.Based on the vehicle hardware platform,the data fusion strategy of multi-sourcesensors is designed.Through mapping from radar coordinate system to spherical coordinate system and mapping from spherical coordinate system to pixel coordinate system of image,the conversion model between radar coordinate system and image pixel coordinate system is constructed,and the radar data points and image pixel points are correlated.A vehicle front-end traffic scene recognition model based on semantic segmentation is constructed.The model improves the structure of Deep Lab V3 Plus network by establishing the transformation relationship between space and depth.The relationship between primary features and decoding module in model coding module is established.Deep-to-space conversion replaces the up-sampling process and enhances the output details.High-level decision information is obtained by fusing radar data and segmented scene target information.The test results of the method on International Open datasets are better.The testing IOU is larger than 0.8.3)A lane marker recognition algorithm based on Markov chain is proposed.Firstly,the image is divided into different sub-regions.According to the image entropy distribution,the image is divided into Lane area and invalid area.Then the image preprocessing is carried out according to the lane features,and the edge feature image of the lane is obtained.Hough is used to recognize the straight line in the image,considering the position and slope of the lane line at the current time,and combining the lane state at the front and back time.The state transition probability of the lane line is calculated by using Markov chain.The maximum transfer probability is selected as the result of lane marker recognition.The algorithm is validated by using the domestic road scene data set,and the accuracy of lane recognition reaches 94.8%.4)A vehicle driving risk assessment method based on fuzzy reasoning is proposed.In order to study the relationship between driving behavior choice and vehicle driving risk under different driving conditions,based on road scene type recognition,road scene segmentation and lane detection in Chapter 2 to Chapter 4,and taking driving behavior selection as a bridge,the driving risk characteristic parameters which can well represent free flow state,car-following state and lane-changing state are selected.Based on the fuzzy reasoning technology,the membership function of characteristic parameters and the reasoning rules of driving risk in free flow state,car-following state and lane-changing state are designed.A real vehicle test platform is built to test the application effect of the evaluation method proposed in this paper based on the road traffic scene information identified and the vehicle running state data in the natural driving process.