Research On Lane Line And Drivable Area Detection For Intelligent Perceptio

With the rapid development of deep learning and the automotive industry,driverless cars and vehicle-assisted driving systems have become an emerging research area of great interest.In the process of vehicle driving,the environmental perception system is responsible for collecting external road information and providing timely feedback of detection results to support the decision making of driverless cars and vehicle-assisted driving systems.Therefore,the performance of the environmental cognitive system is critical to the safe driving of the vehicle.The visual cognitive system is an essential part of the environmental cognitive system,among which lane line detection and driveable area detection are the core technologies of the visual cognitive system for driverless and vehicle assisted driving systems.Therefore,the research on lane detection and driveable area detection is crucial to improving vehicle safety while driving and also provides significant support for the growth of intelligent transportation.In this paper,the limitations of traditional lane detection and drivable area detection methods are studied,and a semantic segmentation-based lane detection and driveable area detection algorithm is presented.Although the semantic segmentation method has certain advantages,there are still problems such as poor real-time performance and complex models.To solve these problems,the following research work is made in this paper:To begin with,a multilevel feature fusion lane detection algorithm based on atrous spatial pyramid pooling is proposed for the lane detection task.The proposed algorithm adds the newly proposed ASPP-Plus module to the codec for multi-scale feature aggregation,thus enabling the network to learn global feature information and solving the problem of no visual cues in lane detection.In addition,the proposed algorithm also presents a multi-level feature fusion decoder for fusing multi-scale upsampled features,which solves the problem of lane edge information loss.Next,a driveable area detection algorithm based on an improved Deep Labv3+ is proposed for the driveable area detection task.To improve the detection efficiency,the proposed algorithm substitutes the backbone for Mobile Netv2 in the network,thus dramatically reducing the number of parameters in the network.However,this also brings the problem of accuracy degradation.To solve this problem,the proposed algorithm introduces a multi-head self-attention mechanism and CBAM attention mechanism to enhance the perceptual field and feature representation of the network,thus improving the detection accuracy of the network.At last,the lane detection algorithm and the driveable area detection algorithm were trained and tested on the public dataset and achieved competitive detection results.The lane detection algorithm achieved 73.2% and 96.49% detection accuracy on the CULane dataset and the Tusimple dataset,respectively,and was able to meet the real-time requirements.Meanwhile,the driveable area detection algorithm obtained 84.2% detection accuracy on the BDD100 K dataset,which also met the real-time detection requirements.In order to better evaluate the segmentation performance,the lane detection and drivable area detection results were also visualized,and the results showed that the proposed algorithm performed well in various scenarios.