3D OBJECT DETECTION from MOBILE LIDAR POINT CLOUD with DEEP LEARNING

The usage area of LiDAR technology, which can be detected from the aerial, terrestrial and mobile, is expanding day by day. Especially for mapping and autonomous vehicles, mobile LiDAR offers very useful data. Mobile point clouds are a type of data obtained using laser scanners mounted on a moving vehicle. An accurate sense of space and precise positioning are crucial requirements for reliable navigation and safe driving of autonomous vehicles in complex dynamic environments. Recently, deep learning approaches have been preferred for the evaluation and information extraction of complex mobile LiDAR data. Although successful results have been obtained for camera-based solutions with deep learning, it may not be fast enough in inference paths due to convolution operations. There are improved methods for real-time performance in object detection. Single-shot detectors, like YOLO, are some of the best in this regard. In this study, moving or stationary vehicles, people and cyclists on the point cloud have been detected by deep learning. Vehicles, pedestrians and cyclists were detected with YOLO3D-YOLOv3 and YOLO3D-YOLOv4, which are the developed version of the YOLO algorithm applied to 2D images for 3D point clouds. KITTI benchmark dataset was used in this study. The point cloud is projected onto a grid mesh with a resolution of 0.1 m per pixel in the form of a bird's eye view. The range of a LiDAR patch is 30.4 meters to right and 30.4 meters to the left, and 60.8 meters forward. Input shape of 608x608 per channel is obtained by using this range with the resolution of 0.1 m per pixel. Average mean precision (mAP) results in this study were obtained within the mAP lower limit of 0.5 IoU for each object class. The mAP was obtained as 83.04%. with YOLO3D-YOLOv4 and 81.50% with YOLO3D-YOLOv3.