Lane Detection Based On Convolutional Neural Network

Lane detection,an important technology of driving assistance system,can effectively prevent vehicles from deviating from the lanes to avoid traffic accidents.Traditional algorithms usually perform inverse perspective mapping on the original image and then detect lanes with Hand-crafted features.With the rapid development of deep learning,researchers have applied deep neural networks to lane detection to deal with the challenges of complex road scenes and different weather conditions in recent years.In this thesis,some problems of lane detection in the original image without inverse perspective mapping are studied,and two algorithms based on convolutional neural network are proposed.Details are as follows:1.Lane detection algorithm based on original image and object detection network(1)The Global Grid Lane Net(GGLNet)is proposed.The detection of thin and long lanes is a problem in object detection network.The feature extraction network ResNet-32 is designed in the GGLNet to increase the number of network layers and improve the learning ability of network.Furthermore,a lane mask detection module,which uses grid-level masks to detect lanes,is developed to improve the accuracy of lane detection.(2)The grid-level masks are difficult to be distinguished and further fitted into different lanes because the lanes in the original image are unparallel and converge to the far-end.In this thesis,a lane location label prediction module,which divides the grid-level masks into different mask sets for different lanes,is designed in the GGLNet.Moreover,a lane fitting module,combining of random sample consensus and least squares method,is further developed to correct lane locations and improve the fitting accuracy.The experimental results on the Chinese Urban Road Dataset,which includes complex road scenes and multiple weathers,show that the proposed algorithm achieves a precision of 98.5%and an accuracy of 88.0%.2.Lane detection algorithm based on original image and semantic segmentation networkThe sample imbalance between lanes and backgrounds is a challenge when semantic segmentation is applied.Based on a lightweight network,ICNet(Image Cascade Network),the weight of sample classes is introduced in the loss function which increases mloU from 51.8%to 67.1%.Furthermore,a lane post-processing module based on K-means clustering and least squares method is implemented to effectively distinguish and fit different lanes.The experimental results on the Baidu Unmanned Vehicle Lane Detection Challenge Dataset show that the proposed algorithm achieves a precision of 93.7%and an accuracy of 92.0%.In conclusion,the two proposed algorithms in this thesis ean effectively detect thin and long lanes,distinguish and fit converged lanes,and reduce the effects of sample imbalance.The methods and the results can be used in future driving assistance systems for better lane detection.