Neural Network-based For Structured Lane Detection Algorithm Research

The transportation industry is the lifeblood of the country’s economy,greatly promoting the development process of China’s economic internationalization and modernization.With the growing demand for intelligence in the transportation industry,artificial intelligence technologies are beginning to be widely used in intelligent traffic.Lane detection is an essential component of automated driving systems,because realtime and accurate capture of lane information on the ground is the key to realize lane drift warning and trajectory planning.Traditional lane detection algorithms based on image processing first extract lane features such as color,texture,and shape of lanes,and then use clustering algorithms or segmentation algorithms to detect lanes.However,traditional detection algorithms rely on hand-crafted lane features,resulting in their ability to detect only lane in fixed scenes.Deep lane detection algorithms based on deep learning have powerful feature learning capabilities and can automatically acquire lane features,thus making them suitable for lane detection in complex scenes.Deep detection algorithms can be divided into image segmentation algorithms and non-image segmentation algorithms.Although all of these algorithms have achieved a very large number of research results and have realized high detection accuracy and fast detection speed in most scenes.However,in complex scenes,such as poor lighting conditions,congested traffic roads,the presence of a large number of targets on the ground with the same characteristics as lane or severe wear and tear of lane information,etc.,lead to the detection performance of these depth algorithms cannot meet the application requirements of practical tasks.Therefore,this paper tries to improve on two existing algorithms to achieve good detection accuracy and detection speed results in complex scenarios.In summary,the research work of this paper is as follows:(1)An efficient lane segmentation algorithm is designed to address the problem of low detection speed of existing image segmentation algorithms.Firstly,by improving the residual networks,this paper proposes a feature extraction module(Spat Res)that can learn both global spatial information and local feature information of the object,and optimize the feature extraction capability of the backbone networks by using the structural characteristics of lane.Then the spatial attention module(SAM)is introduced to calculate spatial attention activation of the output feature map of the decoder in the upsampling process,so that the feature decoder can obtain more accurate upsampling results.Finally,a spanning layer refinement structure(CLD)is designed,which can achieve better pixel segmentation accuracy of the decoder with very low computational cost.The experimental results on Tu Simple and CULane show that this algorithm can achieve very fast detection speed in complex scenes,and meet the accuracy and speed requirements of real-time detection tasks.(2)A high-precision lane detection algorithm is designed to address the problem of low detection accuracy of existing non-image segmentation algorithms.Firstly,a feature distillation network(FDN)is designed to achieve the improvement of network’s feature extraction performance by fusing the multi-scale features extracted by the backbone network at different depths.Secondly,a self-attention distillation mechanism is introduced,aiming to further optimize the target features extracted by the network without increasing the network computational cost and the number of parameters.Finally,a non-local attention mechanism is introduced and the expectation of category prediction is used to replace the category prediction to achieve better detection results.Experimental results show that this algorithm can achieve high detection accuracy,which is well suited to the needs of practical detection tasks.