Global Lane Detection Algorithm Based On Improved CNN In Complex Scenes

With the continuous increase of computing power recent years,the field of computer vision has made significant progress basing on deep learning methods,and the field of lane detection has also developed rapidly with the east wind,although the existing lane detection based on deep learning is far surpasses traditional algorithms in terms of robustness and other aspects,and it’s practical in some simple scenes.However,the uneven distribution of light,traffic,lane wear,night,etc.In urban road scenes such as landmark arrow interference,the result of algorithm is not optimistic.Therefore,this article attempts to improve on three existing lane detection algorithms basing on CNN,which include general image segmentation algorithm(Spatial CNN),classification algorithm(UFLD),which can be used in complex roads,to achieve more accurate detection in the scene.In summary,the research work of this article is as follows:(1)Aiming at the problem that the existing Spatial CNN is not strongly related to global features,a method of lane detection in complex scenes that introduces BRNN and attention mechanism to improve Spatial CNN is proposed.First,select the Resnet-50 convolutional neural network as the basis to build the model,then establish the connection between the slices through the BRNN,conduct the attention mechanism on the two features output by the BRNN to achieve the detection of the lane.Finally,the results of simulation prove the algorithm can detect lane accurately in complex road scenes.(2)Aiming at the problem of insufficient receptive fields in the existing UFLD,a method for lane detection in complex scenes is proposed by introducing supplementary receptive fields and Gaussian heatmap regression to improve UFLD.Firstly,considering the different sizes of the lane features due to the near and far factors in the image,therefore design grids of different sizes to supplement the receptive field,secondly,the attention mechanism is used to obtain the attention value,Finally,fuse the confidence of the Gaussian heatmap and attention value,the results of simulation prove that the algorithm can accurately detect lane accurately in complex road scenes.