An Edge-Server Partitioning Method for 3D LiDAR SLAM on FPGAs

3D LiDAR SLAM (Simultaneous Localization And Mapping) is utilized for autonomous driving and autonomous mobile robots. When compute resources on an edge device are limited, SLAM is executed by using the edge and server cooperatively. In this paper, we propose an edge-server cooperative 3D LiDAR SLAM based on the state-of-the-art method, LIO-SAM. In the edge-server cooperative SLAM, data transfer between them imposes an overhead. We partition the LIO-SAM method to reduce the data transfer while meeting the time constraints; the edge part transfers a set of feature points instead of a raw LiDAR scan, and the server handles the other parts. Moreover, the edge tasks are accelerated by an FPGA implementation to meet an execution time constraint on the edge side. Specifically, we implement a point cloud deskew on the FPGA. Average execution time on the edge side is accelerated up by 1.63 to 1.73 times. Furthermore, all the LiDAR scans are processed within 100ms which is the execution time constraint on the edge side based on 10Hz LiDAR frequency. We also evaluate the data transfer size and the execution time in detail.