Research On Lane Detection Algorithm Based On Deep Learning

Autonomous driving is a hot topic in recent years,lane detection plays a key role in autonomous driving system.Lane detection provides important information for autonomous cars as the guidance of automatic cruise,lane keeping and overtaking on the road.Traditional lane detection algorithms often design hand-craft features based on specific road scene,are hard to handle the challenge of the complex and changeable real road scene environment(such as passing cars,shadow of the building,illumination variety,broken and occlusion lanes),thus get poor robustness.Deep learning is a data-driven technology,has achieved many good results in various fields of computer vision,provides new idea for lane detection research.In this thesis,lane detection based on deep learning is researched.It consists of host lane detection and multi-lane detection.The main work and innovation of this thesis are as follows:1)In the host lane detection algorithm,we use full convolution network model to preprocess the input frame,thus can automatically extract region of interest and fast locate the lane area.We suggest a host lane model and propose host lane feature points extracting algorithm based on hierarchical clustering to effectively remove the disturbance of the noise.2)We make use of the inter-frame similarity of the video sequence and propose a simple and effective host lane correcting algorithm.Experiment results show that it can significantly decrease the false positive rate of host lane detection.3)In the multi-lane detection algorithm,we design a shared encoder and multi-task decoder model.Joint training is carried out on the task of semantic segmentation of lane segmentation and the task of instance segmentation of different lanes,which can improve the training efficiency of neural network.In order to get semantic and instance labels for training model,we develop fine annotation algorithm to annotate coarse labeled dataset.In the instance segmentation decoder,we design loss function based on metric learning,which can distinguish pixels from different lanes.We use morphological operations to postprocess the semantic segmentation results and use clustering algorithm successfully distinguish different lanes in the instance segmentation maps.In this thesis,experiments are carried out on publicly published datasets and self-built challenge road scene dataset,which can demonstrate the effectiveness and practicability of our lane detection algorithm based on deep learning.