Résumé: One of the primary responsibilities of an autonomous driving system is lane detecting. We propose modeling lane markings using Catmull-Rom curves, as opposed to segmentation-based approaches and point detection-based methods, which input images captured by monocular cameras and output lane markings represented by Catmull-Rom segments. The Catmull-Rom segment fits the lane lines more closely than the cubic polynomial and is stable and simple to construct (tested on the TuSimple dataset). We also suggest a spatial attention module to take advantage of lane lines’ nearly vertical distribution. Our suggested approach strikes a balance between accuracy and real-time. In the Tusimple dataset, our model’s accuracy is 0.45% more accurate than the cubic polynomial model (PolyLaneNet). These results demonstrated the suitability of our approach for the lane detecting task.

Langue: Anglais