A curb edge detection method integrating instance segmentation and point cloud analysis

Curb edge detection is a key technology in autonomous driving systems and plays a significant role in path planning and edge-hugging driving. The shape of the curb is usually complex and variable, and in complex scenarios such as uneven lighting and occlusion, there are often situations of missed detection and low accuracy. Based on this, a method integrating instance segmentation and point cloud analysis is proposed. The mask information of the curb is obtained by using the improved instance segmentation model. A feature extraction module combining the convolutional neural network (CNN) and the Transformer attention mechanism is designed to enrich the semantic information of the feature map. An edge-assisted loss function based on the Canny operator is added to improve the segmentation accuracy of the curb in complex scenarios. On this basis, a three-dimensional reconstruction of the curb is carried out, and a method for fitting the curb edge combined with three-dimensional information is proposed. The test results show that the root mean square error (RMSE) of the obtained edge line compared with the manually marked edge line is 3.3, and the processing time of the algorithm is 0.062s. The method proposed in this paper can accurately detect curb edges through experiments in various scenarios.