Lidar Point Cloud Object Detection and Semantic Segmentation Fusion Based on Bird's-Eye-View

Abstract

Self-driving or assisted driving of self-driving cars are based on identifying and categorizing obstructions. This paper mainly intends to create a precise object detection frame based on radar or lidar to collect information and to detect and categorize obstructions. 3D objects in point cloud are often represented using a 3D box, which is similar to how objects are represented on top of a 2D image. But objects in the 3D world do not follow any particular orientation, additionally, box-based detectors can hardly enumerate all directions. So, this model use key point detector to detect the objective's center and return to other attributes, which include the size of 3D box and direction. In the meanwhile, we find that huge discrepancies exist in the 3D box's frontier of pure objective detection and actual objective's frontier. It is hard to directly detect the objective's frontier in 3D point clouds. As a result, this model use the point cloud semantic segmentation method based on Bird'View and polar coordinates, and use the results of semantic segmentation to optimize the boundary of the 3D box. The model in this paper achieves 65.7% and 66.7% mAP for vehicles and pedestrians at level2, with a speed of 19.38 FPS