Video-based Lane Detection For Complex Road Conditions

Vision-based line detection method is a common solution in the industry.The method is based on the image or video algorithm to detect the position of the lane when the vehicle is driving from the picture of the camera.The detection of lane in cities is recognized as a very challenging task.Because in the city,there will be a lot of front cars and pedestrians during the driving process.The types and styles of lane in the city are very different.The wear and lack of lane greatly increase the complexity for this task.In addition the extreme weather such as rain,fog,snow and other extreme weather increased the challenge of this task.Because of this,the task of detecting lane in complex road conditions is an urgent needed.This thesis first summarizes and discusses the research status and progress of lane detection at home and abroad.Through research,it is found that the traditional lane detection method has poor robustness,and the characteristics of the lane are simple and multiple styles.It is difficult to use the traditional method to detect the lane in real life.Moreover,in the field of video lane detection,there are few optimizations and designs of lane detection methods based on video characteristics.Therefore,in view of the above considerations,this thesis has done the following work:1.A three-branch neural network is proposed.First,the three-branch structure is used to extract the lane features at different scales in the picture,and then the network model is compressed by the lightweight module.Finally,the feature recalibration module is used to assign weights to feature channels to improve network accuracy.Training and testing the three-branch network on the complex road condition data set.Experimental results show that the three-branch network can achieve high accuracy on complex road condition.2.By considering the characteristics of video,a lane detection model suitable for processing video is designed.First,the information guidance module is used to pass the result of the previous frame to the lightweight network of the next frame,and then use the cascade method to improve the encoder module for feature extraction capabilities and accelerate network training.Training and testing the video lane detection model on the video data set.Experimental results show that the model can reduce the timeconsumption of the network by 40% while maintaining high accuracy.Finally,the pixel-level semantic segmentation of the lane area in the video is realized.