Accelerating Point Cloud Ground Segmentation: From Mechanical to Solid-State Lidars

TL;DR

This work addresses the need for real-time ground segmentation on solid-state Lidar (SSL) by evaluating point-, grid-, and range-image-based methods, identifying range-image-based RDG as the most robust under frame slicing. It then designs a scalable FPGA-based parallel accelerator, leveraging a five-slice subdivision of SSL frames and the RDG backbone to achieve up to 30.9× speedups over CPU baselines with modest power (≈3.52 W) and a 167.54 MHz target. The authors validate the system on a custom camera-SSL dataset collected on a test vehicle, demonstrating strong IoU/F1 performance (≈89–94) and substantial real-time throughput (≈0.28 ms per frame) while highlighting practical considerations such as edge-edge overlap between subframes. Overall, the study shows that parallel processing and SSL-tailored architectures can deliver efficient, scalable ground segmentation for autonomous perception in edge devices, with clear pathways for deployment and further optimization.

Abstract

In this study, we propose a novel parallel processing method for point cloud ground segmentation, aimed at the technology evolution from mechanical to solid-state Lidar (SSL). We first benchmark point-based, grid-based, and range image-based ground segmentation algorithms using the SemanticKITTI dataset. Our results indicate that the range image-based method offers superior performance and robustness, particularly in resilience to frame slicing. Implementing the proposed algorithm on an FPGA demonstrates significant improvements in processing speed and scalability of resource usage. Additionally, we develop a custom dataset using camera-SSL equipment on our test vehicle to validate the effectiveness of the parallel processing approach for SSL frames in real world, achieving processing rates up to 30.9 times faster than CPU implementations. These findings underscore the potential of parallel processing strategies to enhance Lidar technologies for advanced perception tasks in autonomous vehicles and robotics. The data and code will be available post-publication on our GitHub repository: \url{https://github.com/WPI-APA-Lab/GroundSeg-Solid-State-Lidar-Parallel-Processing}

Figures (7)

• Figure 1: Mechanical Lidar Mechanism and Frame Segmentation
• Figure 2: Top: The demonstrate data frame, captured using the SSL (Robosense M1) facing a flat wall, consists of a total input point sequence of 78,750 points. Upon observation, each subframe is slightly misaligned, with adjacent subframes overlapping at the edges. Additionally, some points are missing due to transmitter or receiver errors, as the Lidar sensor used was trial equipment. Bottom Left: Illustration of the second subframe from the left, showing the uniform rectangular format of all the subframes. Bottom Right: Statistical analysis of values in the x, y, z, and validity dimensions, clearly depicting the subframe’s uniform rectangular shape with dimensions of 126 x 125.
• Figure 3: SSL Lidar Mechanism and Frame Segmentation
• Figure 4: SSL Processing Acceleration Architectures: Implementation across different processing unit allocation strategies
• Figure 5: Resource Usage and Execution Time Across Different Levels of Parallelism: execution time is measured in ms, and resource usage value is presented as normalized values relative to a single processing unit. When the design works at 167.54 MHz, the estimated exuection time for a single PU setup is 0.137ms