Research On Lane Keeping Method Based On End-to-end Learning

Lane keeping is an important component module in autonomous driving systems.This paper studies the direct application of low-cost monocular vision sensors to control the steering angle of smart cars,and then to achieve self-driving.Traditional lane detection methods are susceptible to lighting,occlusion,shadows,and lane line breakage,which pose severe challenges for algorithms based on image segmentation and edge detection,and eventually lead to a serious decline in robustness and generalization of artificially designed feature detection models.In recent years,deep learning has been widely used in various fields because of its powerful ability to approach highly nonlinear functions.In addition,convolutional neural network(CNN)has a prominent advantage in the field of computer vision,which is able to learn an end-to-end depth model from massive image data for extracting key features of lane retention.Therefore,this paper mainly researches lanekeeping methods based on convolutional neural networks.The main work is as follows:The learning process of convolutional neural networks takes a lot of time.In order to solve this problem,we achieve the purpose of reducing the time cost by cutting the original image into a reduced size,but the importance of sky,roadside and road-related features cannot be known.Therefore,we study the performance of end-to-end controllers trained with different feature sets in simple and complex environments.We design three experimental frameworks to study the effect of a single feature end-to-end depth model on vehicle steering control,and provide feature selection criteria to reduce the computational resource cost and time cost while ensuring the good robustness of the model.End-to-end control models based on deep learning are mostly trained by twodimensional convolutional neural networks(2D CNN).Because the inter-frame motion information in the time dimension is not considered,the interpretability and generalization ability of the depth model are poor.The three-dimensional convolutional neural network(3D CNN)can learn spatiotemporal features from continuous video frames.In addition,Deep Deterministic Policy Gradient(DDPG)is also commonly used for control optimization of continuous actions.However,the DDPG algorithm still has the problem of low sample utilization caused by unreasonable sampling methods.Based on this,a method combining 3D CNN and improved DDPG algorithm is used to accurately predict the vehicle steering angle.The current deep learning-based lane keeping programs lack a standard database and are mainly focused on research in the same domain.During the research process,we found that the deep model trained with previously collected data is difficult to extend to subsequent scenarios.Nevertheless,robust perceptual-action models should be learned from data with different scenarios and real behaviors,the current model learning is generally limited to training of massive data,innovation of deep network architecture,and in-situ models learned in the simulation environment.Inspired by fast style transfer and artistic style neural algorithms,we propose an arbitrary style generation network architecture,which includes a style transfer network,a style learning network,a style loss network,and a multivariate Gaussian distribution.The style-embedding vector is randomly sampled from the multivariate Gaussian distribution and linearly interpolated with the embedding vector predicted by the input image on the style learning network,which provides a set of random normalization constants for the style transfer network,and ultimately achieves image style arbitration.Based on this,this paper explores the introduction of the proposed new style transfer method into data augmentation technology,which improves the diversity of limited data by changing the texture,contrast,and color of the image,thereby generalizing it to scenes not previously observed by the model.The proposed algorithm is simulated in Udacity's self-driving simulator,and the intelligent vehicle experimental platform is constructed by using microcomputer Raspberry Pi.Simulation and real vehicle experiments show that our proposed method has good generalization and robustness,and can maintain a high lane-keeping success rate in different scenarios.