Research On Adaptive Lane Departure Warning System Based On Monocular Camera

Road traffic safety has received increasing attention in recent years.According to Ministry of Transport of the People’s Republic of China,approximately 50% of car accidents were caused by vehicles deviating from lanes.The Federal Highway Administration indicated that 44% of all fatal traffic accidents in the United States involved lane departure in 2002,which was also considered to be the significant cause of vehicle rollover accidents.Lane Departure Warning System(LDWS),as an important innovative technology to improve road safety,has reduced the casualty rate by more than 50% and effectively avoided the occurrence of traffic accidents.However,there still exists a challenge that the accuracy of early warning is greatly affected by environmental light condition.LDWS consists of lane line detection model and lane departure warning model.In order to improve the accuracy and robustness of the lane departure warning system,I deeply study the lane detection model and the departure warning model.Then,on this basis,I practice the lane departure warning system on the experimental vehicle.The main contents of this thesis are as follows:This thesis proposes a lane detection method based on deep learning.According to the different environmental light condition in the images captured by the vehicle,the images are roughly divided into six categories: day,night,backlit day,backlit night,shadow,and backlit shadow.I propose a model for light condition classification which is a residual network that fused with self-attention mechanism.Different results of classification correspond to different candidate lane detection models which complete the lane detection task.The candidate lane detection models in this thesis are the PINet model and the Lane ATT model.After completing the detection of the key points of the lane,I transformed the image into a top view using the image inverse perspective transformation method,and use the third-degree polynomial to fit the lane line.In order to reduce the system’s misjudgment due to the missed detection of lanes in a single frame,I apply Kalman filter to track the fitting curve to ensure that the lane detection results are consistent in the time sequence.About lane departure warning,I first analyze and compare different departure warning models and their advantages and disadvantages.On the basis of considering the direction and speed of the vehicle,I also considered the yaw angle,and finally chose the Time to Lane Crossing(TLC)early warning model as the system’s departure early warning model to reduce the false alarm rate to a certain extent.The complete lane departure warning system is built on a experimental vehicle which based on the YHS FR-09 drive-by-wire chassis and equipped with a PC based on Ubuntu 20.04.Finally,the system is fully tested in the actual environment.Experiment results indicate that the system is effective and the accuracy and robustness expected of the lane departure warning system have been achieved.