LaneGuardMapper: A Framework for Real-Time Lane Detection and Tracking in Challenging Driving Conditions

Engineering notes

Key topics: autonomous driving, autonomous vehicle, lane detection, semantic segmentation. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

The progress of autonomous driving and driver-assistance systems greatly depends on having reliable lane detection systems. A new approach, LaneGuardMapper, is suggested in this paper to detect and track lanes in real time, no matter the driving circumstances. It relies on edge segmentation, time consistency and changes to suit different environments for maximum accuracy and stability. Using the TuSimple dataset and by adding night, rain and fog situations that are created by GAN, the model can detect lanes even when the vision is greatly affected. Semantic segmentation relies on an Spatial Convolutional Neural Network (SCNN) which is made more effective by adding attention and dilated convolutions for straight lane detection. Using 3D CNNs and optical flow fusion helps with temporal consistency and gradient lane fitting together with Kalman smoothing are used for reliably tracking the vehicle. The system reached an accuracy level of 95.93% in all the 10 scenarios which is more than 5% greater than what traditional models can achieve. The car managed to achieve 93.2% accuracy and 94.2% recall while it was foggy and at night. Experimentation with Jetson Xavier NX and Jetson Nano proved that the system can run faster than 30 ms in real time. The demanding situations in road scenes such as lane merging, covered objects and difficult lighting, do not prevent the framework from adapting. LaneGuardMapper does a good job at identifying and tracking lanes which makes it suitable for future autonomous vehicles and helps with intelligent transportation. By using multi-sensor data fusion, applying training in other domains and increasing continual learning, future progress can improve performance more.

Links and sources

• Official / arXiv page

Comments