Research On Fast Response Lane Detection Based On Semantic Segmentation And Optical Flow Estimation

In recent years,with unceasing development of global economy,there are more and more vehicles on the roads.Consequently,urban traffic jams and accidents have been increasing.Advanced driving assistance systems(ADAS)and unmanned driving technology can reduce accidents and improve transportation efficiency by reminding the driver or taking over the operation of the driver.Lane recognition is an important part of ADAS and unmanned driving technology,which can provide feedback for lane departure warning,path planning,and lateral control,etc.However,fast lane recognition algorithms are easily affected by illumination conditions,occlusion,weather,etc.High-precision lane recognition algorithms have a relatively slow recognition speed due to the large amount of computation.Therefore,the research on lane recognition algorithm with robustness and rapidity is the current research hotspot.To solve the above problem,this thesis proposes a fast response lane recognition method based on deep learning semantic segmentation and optical flow estimation.First,the position features of the lane pixels in the image are obtained by lane segmentation.Second,the lanes are classified and denoised according to the position features of these pixels.Finally,the quadratic equation of all the lanes in camera coordinate system is obtained by a monocular vision ranging algorithm and least square method.The specific work mainly includes the following two aspects:In the phase of lane segmentation,video frames are defined as key frames or non-key frames.The lane segmentation of key frames is completed by the semantic segmentation network;the lane segmentation of non-key frames is completed by the optical flow estimation network and key frame segmentation results.An extremely small neural network makes decisions about the property of the frames based on the underlying optical flow features of the video frame.In the phase of post-processing,first,the lane pixel position feature in the lane segmentation results is used to classify the lane for obtaining the pixel sets of each lane.Second,a monocular vision ranging algorithm is used to map the pixels of the lane from the pixel coordinate system to the camera coordinate system.According to the mapped pixel set,curve is fitted by the least squares method to provide feedback and curvature for lane departure warning and unmanned driving lateral control.The above research is conducted on the open and self-collected data sets.The result of it demonstrates that the proposed method can achieve lane recognition in a variety of scenarios.Compared with the lane segmentation method which simply uses the highprecision semantic segmentation model,the speed of proposed method improves 3 times.