Research On Lane Detection And Classification Method Based On Multi-Scale Feature Fusion

Lane serves as a fundamental traffic sign on the highway,primarily tasked with regulating vehicular positioning to ensure the orderly and fluid movement of traffic.With the advent of traffic digitalization and intelligentization,precise identification and localization of lanes not only facilitate vehicles to travel accurately but also have significant research implications for advancing intelligent transportation and smart cities.Mainstream lane recognition methods commonly approach the recognition process as an instance segmentation task.However,they face difficulties in distinguishing varying numbers of lanes found in actual road scenes,requiring the pre-defining of maximum lanes to distinguish different lane instances,which is proved inadequate in complex road environments.Moreover,lane detection is a global super-sparse task.During the lane recognition process,high-level semantic features provide more abstract lane information,whereas low-level image features deliver more precise lane position information.Therefore,integrating high-level and low-level features can highly complement the lane recognition process.However,current lane recognition methods do not optimize this integration effectively.Although the row classification methods have bridged the gap by resolving the issue of high resource utilization tasks,it still lacks the seamless integration of features of different levels.This paper undertakes two aspects of research,namely:(1)To address the issue of inaccurate lane alignment arising from inadequate feature integration at different levels in current lane alignment detection methods,this paper proposed a lane alignment detection method that integrates multi-scale contextual information.By leveraging this method,location information in low-level features can be effectively transferred to high-level features,significantly enhancing the semantic expression ability of high-level features.Experimental results reveal that,compared with existing methods,the F1 score of the lane detection method that utilizes multi-scale contextual information is improved by approximately 2.4% and 3.8% on the Tu Simple and CULane datasets,respectively.Furthermore,the proposed method reduces the required time by approximately 4.5ms and3.7ms on the respective datasets while maintaining high real-time performance.(2)To overcome the limitation of the maximum number of lanes set in lane recognition methods based on instance segmentation that poses challenges in distinguishing different lane instances,this paper presents a lane instance classification method based on neighbor inner product.Initially,lanes from images are extracted and transformed into discrete coordinate points.The subsequent classification of lane coordinate points is determined by constructing vectors between adjacent coordinate points for inner product operation.Different lans are ascertained by the discrete coordinate points.Experimental results reveal that the proposed method outperforms alternative methods,with the F1 score increasing by up to 4.75% and with higher precision and recall levels of 95.35% and 95.28%,respectively.