Research On Lane Detection Algorithm In Abnormal Driving Environment

With the rapid development of economy,vehicles have become the preferred choice for people to travel,but the popularity of vehicles has also led to the occurrence of many traffic accidents,most of them occur in the entrance and exit tunnel light and dark changes,the high beam on the vehicles at night and other abnormal driving environment.As an important safety constraint for vehicles to drive according to the established route,lane detection in unusual driving environment is of significance to ensure driving safety,however,there is little research on lane detection technology in abnormal environment.In this paper,we propose four end-to-end lane detection algorithms in abnormal driving environments based on computer vision using deep convolutional neural network.(1)In terms of algorithm,we propose a lane detection algorithm based on semantically segmented ENet,compared with the traditional lane detection algorithm.ENet is created and trained,and the experimental results are analyzed and visualized.The performance of the test is 74.8%for mIoU and 83.6%for mPA,17.2 ms for inference time,58 FPS,and 4.1 M.Experiments show that ENet has good detection performance for simple road environment scenarios,but its detection capability is weak for abnormal driving environment.For this reason,we design three improved end-to-end lane detection algorithms based on the basic ENet lane detection algorithm to improve the accuracy and real-time goals of the lane detection task:1)To meet the requirement of real-time detection,a lane detection algorithm O-ENet based on object-decoder distillation is proposed,which reduces the model complexity and computational overhead by means of model compression,and the algorithm can be extended to any semantic segmentation network using an encoder-decoder architecture.The performance on the test is 72.9%for mIoU and 80.4%for mPA,with an inference time of 11.6ms and an FPS of 86,and model parameter number of 1.9M,which demonstrates that the algorithm can detect lanes in abnormal driving environments more quickly with guaranteed detection accuracy.2)To meet the requirement of accurate detection,a lane detection algorithm A-ENet based on multi-attention perception is proposed,compared with other deep learningbased lane detection algorithms,the algorithm adopts a new perspective from the perspective of road scene understanding and uses an attention mechanism to aggregate global context information.The performance on the test is 80.2%for mIoU and 89.7%for mPA,with an inference time of 29.9ms and an FPS of 33,and model parameter number of 33.4M.The algorithm is shown to be more accurate in detecting lanes in abnormal driving environments while maintaining real-time detection.3)To balance the real-time performance and accuracy of detection,a multi-attention object-decoding distillation composite lane detection algorithm AO-ENet is proposed,combining the advantages of A-ENet and O-ENet algorithms,which performs on the test with mIoU and mPA of 79.7%and 89.2%respectively,with inference time of 19.1ms,FPS of 52,and model parameter number of 7.6M.The experiments demonstrate that the algorithm achieves a good balance between accuracy and realtime performance in lane detection tasks.The optimal lane detection algorithm in this paper is AO-ENet by weighing accuracy and real-time performance as well as the demonstration of relevant experiments.(2)In terms of data,the dataset of lanes in abnormal driving environment required by this project is made,with a total of 6,000 road scene pictures.At the same time,in order to further improve the generalization ability of network model,data preprocessing and data enhancement are used to enrich the dataset samples.In addition,considering the serious class imbalance problem of lane and background in road images,we improve the performance of the network model by the loss function without changing the network structure.Experiments prove that the weighted crossentropy loss function can improve the performance of the algorithm’s detection very well.This article has 63 figures,13 tables,and 91 references.