Research On Key Technologies Of Vision-based Vehicle Driving Assistance System

In order to improve the driving safety of vehicles and effectively avoid traffic accidents,vehicle driving assistance system has become a research hotspot in the field of intelligent transportation.However,the current vehicle driving assistance system cannot provide reliable driving safety guarantee under the conditions of low light,fuzzy and damaged lane,unclear road shape,complex vehicle environment obstacles and various factors interfering with the driver's operation.In view of the above factors that affect the work of vehicle driving assistance system,this paper focuses on the recognition method of structured and unstructured mixed road perception in complex multi-interference environment based on vision,and omni-directional obstacle detection in complex vehicle environment based on depth camera,and driving behavior detection method based on depth camera.The main work includes the following four aspects:(1)To solve the problem of road shadow,rain,stain and reflected light that affect the structured road lane recognition and the problem that lane keeping function cannot work reliably on unstructured road without lanes,unstructured road with fuzzy lane boundary and with lots of interference factors around the scene,inverse projection transform(IPM)and adaptive lane edge extraction method based on improved K-means iterative clustering segmentation and a method of lane edge region search and statistics are proposed,at last using the polynomial function to fit the main control points on lane edges to generate the final lanes.At the same time,an unstructured road extraction algorithm based on the mixture of RGB entropy and Maximum Two-dimensional entropy is given to extract the lane,then using the improved fuzzy entropy algorithm to judge the uncertain region of the road image.At the same time,an improved region growing method is used to search the region of the mixed entropy image and using an improved least square and quadratic curve model to generate the lane boundary.At last,an mixed warning model of lane departure based on improved CCP is established to judge and warn when the vehicle deviates from the lane in mixed road.The experimental results show that the proposed method can effectively improve the detection and perception ability of lane and lane boundaries of mixed roads indifferent interference environments,and ensure the real-time detection and perception speed of mixed roads at various speeds,which meets the real-time requirements of vehicle driving assistance system(2)To solve the problem of narrow detection range,high cost,time-consuming calculation and easy to be affected by environmental light which caused by traditional vehicle environmental obstacle detection method,this paper uses the depth camera to detect omni-directional obstacles in vehicle environment.After determining the ROI area and using the depth camera transformation matrix to generate the panoramic depth image,an obstacle extraction method based on the spatial region growth of depth image and the improved adaptive DBSCAN algorithm is given,using fast depth estimation and inpainting of depth image holes based on improved KNN and improved iterative Normalized Cut segmentation to achieve hole inpainting of depth image and clustering and merging the fragmentary and irregular obstacles.At last,the omni-directional obstacle distribution map of vehicle environment is obtained by inverse transformation matrix of depth camera.The experimental results show that compared with other methods,our method can more effectively and completely detect obstacles in the complex environment during the day and night,and the detection accuracy,precision and recall rate are better.The above advantages provide a strong guarantee for omni-directional obstacle detection by vehicle driving assistance system in vehicle environment.(3)To solve the problem that current vehicle driver assistance system has high cost,reconstruction of vehicle structure,impact on normal driving by detecting the driver's mental state,vehicle driving state inside and outside the vehicle through a large number of sensors installed in the vehicle,this paper uses the depth camera to detect the driver's dangerous behaviors.An improved one dimensional Hough voting based on inverse mapping space growth is proposed to detect and locate the steering wheel in depth vision space,and using the improved Zhang-Suen skeleton thinning algorithm based on hybrid adaptive DBSCAN and FAST skeleton joint extraction and location algorithm based on cluster analysis of multi-mapping connected domain analysis to determinethe skeleton andjoint points of drivers in the depth vision space.At last,the dangerous driving behavior detection method based on decision tree and the dangerous driving state evaluation method based on skeleton entropy are designed.The experimental results show that our method realizes the contactless detection of driver's dangerous behavior,and realizes the ability of all-weather fast detection.The calculation of skeleton entropy comprehensively evaluates the chaotic degree of driving behavior,which is of great significance to prevent the further development of dangerous driving behavior.(4)The software and hardware test platform of the vehicle driving assistance system using our method is built,and the requirements of the vehicle driving assistance system and the software and hardware process of the test platform are described.Based on the campus of Chang'an University and the automobile test field,the joint tests are carried out on the mixed lane perception of structured and unstructured roads,omni-directional obstacle recognition of vehicle environment and detection of driver's dangerous behavior.Through the real vehicle test of this method,the results show that the vision-based vehicle driving assistance system based on our method has strong feasibility,high reliability,strong effectiveness and high efficiency,effectively improves the driving safety,and provides a theoretical basis for the future research of intelligent vehicles.