Research On Lane Detection Method Based On Semantic Segmentation Network

Advanced assisted driving systems play an important role in reducing the occurrence of traffic accidents and improving driving safety,and can effectively improve the increasingly serious traffic problems,and have been widely used in the field of intelligent vehicles.As a means of environment awareness for intelligent vehicles,lane detection technology provides road lane information for assisted driving functions such as path planning,lane keeping and lane departure warning,and is one of the core technologies of advanced assisted driving systems.Traditional lane detection methods rely on manually designed templates to extract lane features,which require good feature matching and cannot adapt to scenes with changing lighting conditions and have poor robustness.The lane detection method based on deep learning extracts features through deep neural networks with better generalization capability,but usually requires predefined number of lane to be detected,which is less accurate in unstructured road scenarios,and the overall method is computationally intensive,less efficient to run in embedded devices,and cannot meet the real-time demand of detection algorithms for smart cars.In response to the above problems,this paper carries out the research of lane detection method based on semantic segmentation network,and completes the fast and accurate detection of lane through the combination of post-processing algorithms such as semantic segmentation network lane extraction and clustering operations,as follows.1.Proposed an improved semantic segmentation based deep neural network lane extraction methodIn this paper,based on the two-branch network structure,a specific structure is designed for the lane shape characteristics of the network to be used for segmenting the lane pixels and background pixels in the image.At the same time,the semantic segmentation dataset of lane in different scenes is established for the network training by data acquisition and labeling.In addition,for the lane detection task,a dynamic region of interest-based road image preprocessing method is proposed to reduce the amount of input data processing and improve the generalization ability of the network for lane detection.The experimental results show that the network can extract the lane in the image more efficiently than other semantic segmentation networks,and has good segmentation accuracy.2.Proposed correlation post-processing method for lane detectionThe lane segmentation map extracted by the semantic segmentation network is postprocessed by the correlation algorithm to build a complete lane detection system.First,a lane clustering algorithm based on lane information constraints is proposed.The proposed algorithm effectively utilizes the constraints of fixed lane interval distance and similar slope between segments to realize the clustering of lane pixel points in the segmentation map separately.Secondly,the problem that the lane model is easily disturbed by noise points when fitted using the least squares method is improved by segment sampling the point set obtained by clustering each lane.Finally,a lane smoothing method based on inter-frame similarity is proposed for smoothing the missed and false detection results in the detection results to improve the accuracy of the detection algorithm.3.Build an experimental platform to verify the algorithmThe lane detection method is experimentally verified through simulation and real car experiments.First,the lane detection system built in this paper is simulated and compared on public and self-built datasets on the computer side.Secondly,the algorithm of this paper is ported to an embedded device and a real car experiment platform is built for real scenario testing.The experimental results show that the proposed lane detection algorithm has good real-time performance and robustness in the integrated road scenario.In summary,this paper presents a semantic segmentation network and related postprocessing algorithms for the lane detection task,and proposes a lane detection method based on semantic segmentation network for experimental validation.The proposed method extracts the lane by semantic segmentation network and combines with post-processing algorithms such as lane clustering to achieve fast and accurate detection of lane,which has certain practical significance and engineering application value.