Lane-Crossing Detection Method Of Vehicles Based On Vehicle Image

In recent years,with the increasing number of vehicles,the negative impact of bad and illegal driving behaviors on traffic system is becoming more and more serious.Among these behaviors,lane-crossing is a common and dangerous one,which often leads to illegal lane-changing,aggravated traffic congestion,traffic accidents and other consequences.Detection of lane-crossing behavior can help automatically recording violations,detering traffic offenders,optimizing traffic flow distribution,assisting driving and so forth.Most of the existing researches use surveillance cameras to detect line-crossing behaviors in specific areas.However,the cost of surveillance cameras is higher and the coverage is much smaller than that of vehicle cameras.To tackle this issue,this dissertation focuses on the research of lane-crossing detection method based on vehicle image,including the lane-crossing detection of the target vehicle in front and the vehicle itself.First,supervised learning based methods require a lot of annotated data for model training,but the acquisition of real lane-crossing detection data comes with many problems,such as the environmental conditions are difficult to meet,the security of the acquisition process,and the annotation of data consumes much manpower and many material resources.Regarding this issue,we use Unreal Engine 4 to build several realistic traffic scenes,in which we use vehicle cameras to get picture sequences and collect data with help of the plugin Airsim and physics engine.The data annotation tool is developed to automatically annotate the collected data,so as to obtain a rich,diverse,well-annotated dataset of lane-crossing detection.Then,combined with image semantics segmentation method,vehicle detection and lane detection are completed,the results of which are given in the manner of segmentation images.Further more,for target vehicle,the positions of front and rear wheels are obtained using the method of front and rear wheel estimation,and the lane-crossing judgment of the target vehicle is realized by comparing the wheel position with the outlines of lane line.For the vehicle itself,geometric parameters such as the inclination angle of the lane lines on both sides of the current lane are detected,by method of the inclination angle double threshold comparison,the lane-crossing behavior of the vehicle itself is judged.Meanwhile,the position and orientation of the vehicle itself in current lane are provided for driver to correct the path.In order to validate the effectiveness of proposed vehicle image based lane-crossing detection method of target vehicle and the vehicle itself,in this dissertation,we conduct comparative experiments on various weather and illumination conditions on the constructed dataset and analyses their effects.Results show that,by using ground truth segmentation images from dataset,we achieve an average precision of 94.4%and 95.7%respectively.With segmentation images from semantic segmentation model,we achieve an average precision of 88.7%and 95.2%respectively,and the average detection time of one single image is 35 ms and 57 ms,respectively.In summary,we start from two aspects,namely data and methodology to work on the research of lane-crossing detection of vehicles.From the data perspective,we use synthetic-data method to build a lane-crossing dataset,and thus remedy the problem of acquiring lane-crossing dataset.From the methodological perspective,we exploit vehicle image based lane-crossing detection methods,and deal with the major difficulty of the detection:getting wheels positions of target vehicle and the double threshold comparison based judgment principle,leading to rather accurate lane-crossing detections.Experiments show that the proposed methods have both high accuracy and good real-time performance,which means they have certain practical values.