Lane Detection Based On Convolutional Neural Network

With the development of society,the traffic safety problem caused by the increase in vehicles has become more and more serious,and people urgently hope that vehicles have the function of autonomous driving.The lane defines the boundary of driving area,and accurate lane detection is one of the key technologies to realize automatic driving.However,due to the complex road environment,lane detection is susceptible to weather,obstructions,and vehicle congestion,and there are still great challenges.There are currently two mainstream lane detection schemes,the traditional method using image processing techniques and the method based on convolutional neural networks.The traditional lane detection method uses basic features such as lane color,shape,and gray level changes to detect lanes,and its accuracy and robustness need to be further improved.The lane detection method based on deep convolutional neural network has better accuracy and robustness,but most of the existing models are too complex and the model parameters are too large for in-vehicle embedded systems with limited hardware computing power and limited storage space,and it is often difficult to meet the actual application requirements in terms of the real-time performance of the mobile terminal.To address the above problems,this paper mainly studies the lightweight lane detection method based on convolutional neural networks.By inserting the preferred attention module into the real-time semantic segmentation network,the lane detection model can meet the real-time requirements of lane detection.The main research work is as follows:(1)To decrease the complexity of lane detection model,the lightweight semantic segmentation-based model ENet and ERFNet are proposed as the backbone network of the lane detection network.(2)This paper uses the self-attention distillation method to solve the problem of sparse lane features by introducing self-attentive distillation in the encoder stage of the model and improves the feature extraction ability of the model through the learning of the attention map between layers.(3)This paper designs a spatial self-attention module to form self-attention and do pooling and convolution in spatial direction to construct spatial attention.The module uses the structure of residual connections,which can be added to any position of the network.After multiple sets of experiments,it is proved that this module can effectively improve the accuracy of the model.(4)In order to increase the model’s global perception of lane and understanding of background semantics,this paper adds a lane classification network to the end of the encoder,adding an additional global supervision signal to the model.This paper uses the CULane dataset to test and verify the proposed lightweight lane detection network.The experimental results show that the lightweight lane detection method proposed in this paper can achieve high accuracy in a variety of scenarios and can meet real-time requirements for lane detection.