Research On Detection And Classification Of Lane Line Based On Deep Learning

Lane-line detection algorithm is susceptible to environmental factors such as illumination and obstructions.To solve these problems,this thesis proposes a lane detection algorithm based on multi-task convolutional neural networks.Through changing the structure of reference network,the detection result is improved.Through compressing network,the running speed is improved.The proposed network can run 60 FPS on PC,which can also run 5 FPS on mobile platform TX2.First,this thesis designs hardware platform and manually makes dataset for training network.This thesis manually annotates 6000 campus lane lines and uses script functions to process them with public data sets to obtain the dataset needed to train the network.Extraction of regions of interest,image scaling,flipping and other operations further optimize and enrich the dataset.Second,this thesis changes the structure of network and compresses it.The network in this article draws on the network architecture of DriveNet and VPGNet.Based on the existing multi-tasking network,this thesis also makes the following two contributions.First,this thesis changes the base network,which is very important for multi-task learning.Replacing the Alex Net network in the original multitasking network with ZFNet and VGGNet networks,the result turns out better.Second,this thesis compress the network.In order to be able to deploy the network on the TX2 platform,this thesis reduces the network model to one-sixth by compressing operations,and increases the operating speed on the PC side and on the TX2 platform by 3 and 3.5 times..Then,this thesis uses distance-based clustering algorithm to cluster the lane-line points,to fit the result by using least square algorithm.In order to obtain pixel-level lane-line detection results,making post-processing operations.Through improving sampling,inverse perspective transform,and clustering,accelerating the post-processing speed.Finally,this thesis deploys a PC-trained network on the TX2 mobile platform.The compressed network can run 5 FPS on TX2 platform.