Road Feature Extraction Based On Deep Learning

With the rapid development of the automobile industry and the rapid increase of the number of cars,people pay more and more attention to the problem of traffic safety.Using artificial intelligence technology to solve the problem of urban road traffic safety is the best method at present.Using intelligent algorithms,we can efficiently process an image,accurately find the road feature objects we are interested in from the image,such as lane lines,pedestrians,vehicles,etc.,and classify the feature objects,thus providing a reliable solution for intelligent assisted driving.However,due to the complexity of the actual scene,the objects to be detected in the road image have problems such as shape damage,uneven illumination,shadow occlusion,and low resolution.How to extract road features more effectively has become a research difficulty.The deep learning method can extract the feature information of the road surface image by itself,which is more discriminative and robust than the features designed manually by the traditional method,and can adapt to the changing environment in the road features.In addition,the diversity of samples makes the trained network more accurate,which greatly surpasses the traditional methods in performance.Therefore,this paper investigates the task of road feature extraction in conjunction with deep learning algorithm,and mainly makes the following work and contributions:(1)Aiming at the problems of shadow occlusion and uneven illumination in lane detection,a lane line detection method based on Lanenet and image enhancement technology is constructed.Firstly,the multi-scale Retinex algorithm is used for color enhancement and noise reduction of the input image.Then,a Bilateral Multi-scale Fusion Network(BMFNet)is designed to realize the information interaction between shallow features and deep features,and capture the context semantics.A new Asymmetric Convolution Pyramid module(ACP)is proposed,which integrates asymmetric convolution into the empty convolution layer with different expansion rates to improve the feature extraction ability of the network and reduce the amount of computation.(2)Aiming at the problem that the proportion of daytime images and nighttime images in the existing lane line data set is extremely unbalanced,a lane line detection method based on generative adversarial network and attention mechanism is proposed.This method uses the generative adversarial network for style transfer to expand training samples.At the same time,a Coordinated Attention Detection Network(CADNet)is designed.By adding an attention mechanism to the residual module of the backbone network,the model learns the correlation between feature channels,automatically calibrates the attention on the channel dimension,and improves the robustness and generalization ability of the model.This paper has carried out experiments on mainstream datasets and compared them with excellent algorithms in recent years.The experimental results show that the method proposed in this paper can reduce the impact of light and shadow on the image detection effect,can detect lane lines more effectively in complex scenes such as occlusion and darkness,with higher accuracy,lower false detection rate and missing detection rate,and can effectively improve the feature extraction performance.