Lane Detection Algorithm Based On Deep Learning

With the development of intelligent transportation,intelligent driving assistance technology has become a research hotspot in recent years,and lane detection is a key technology of intelligent driving assistance.Accurately detecting the position of lane can provide important information for the safe driving of vehicles,so it is of great practical significance to study lane detection.At present,many traditional lane detection methods carry out lane detection through artificially designed features,while such methods are only applicable to a single scene.In the actual driving scene,under the influence of weather environment,lighting conditions,road complexity,lane wear,lane occlusion,etc.,the traditional methods have the problem of high false detection rate and missed detection rate.Neither accuracy nor robustness meets the demands of autonomous driving.With the development of computer technology and neural network,deep learning method has achieved good results in image recognition,computer vision and other fields,and is also widely used in lane detection tasks.Therefore,this thesis adopts deep learning method to study lane detection.In order to improve the speed and accuracy of lane detection,two corresponding lane detection methods are proposed for image-level dataset and video-level dataset respectively:(1)In view of the problem of high computational cost and timeconsuming by current mainstream image semantic segmentation methods,an Attention-based Row Selecting Networks(ARS-Net)image lane detection algorithm combining attention mechanism and row selecting method is proposed.Lane detection is regarded as a row selecting problem,only the location of lane at predefined rows of the image is selected,which greatly reduces the calculation cost compared with the pixel-by-pixel segmentation method.At the same time,the channel and spatial attention mechanisms are used to adjust the weight of features,so as to reduce information loss by paying more attention to important features and suppress unnecessary ones.Meanwhile,group normalization(GN)is used to replace batch normalization(BN),which avoids the influence of batch size and effectively improves the stability of precision.We perform extensive ablation experiments on public image-level datasets Tu Simple and CULane and compare with other stateof-the-art methods,which verifies that the proposed method can achieve the fastest detection speed without reducing the accuracy.(2)Aiming at the problem that the image lane detection method does not take into account the sequence information of the video frames,an improved Multi-level Memory Aggregation Network(MMA-Net)video lane detection method is proposed.The gated recurrent unit(GRU)is used as the encoder,which can obtain long-term memory information better than convolutional neural network.Meanwhile,the number of encoder layers is increased,and the original symmetric structure of MMA-Net is changed to the asymmetric network structure with more encoders and less decoders,so as to extract the deep features of the lanes and obtain the elongated shape structure better.The two-stage training method is used for training,and the attention mechanism is used to integrate the local and global memory features.In this way,the time continuity information of adjacent video frames is effectively used,which solved the problems of lane blur,occlusion and wear,and greatly improves the detection accuracy.We conduct ablation and comparison experiments on video-level dataset VIL-100 and demonstrate that the improved MMA-Net algorithm has significantly better detection accuracy than other methods.In summary,in view of the problems existing in the current mainstream lane detection methods,this thesis designs the lane detection methods in images and videos,which can not only meet the real-time requirements of autonomous driving,but also accurately detect lanes in a variety of complex environments,with high application value.