Real-Time LiDAR Point-Cloud Moving Object Segmentation for Autonomous Driving.

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Artificial Intelligence

Background:

• Autonomous navigation relies on accurate object recognition and prediction for safety and efficiency.
• Existing 3D LiDAR point-cloud moving object segmentation (MOS) models are computationally intensive, hindering real-time performance on embedded systems.

Purpose of the Study:

• To develop a lightweight and efficient MOS network for real-time processing of 3D LiDAR point clouds.
• To enable accurate moving object segmentation on resource-constrained embedded platforms for autonomous vehicles.

Main Methods:

• Proposed a novel lightweight convolutional neural network (CNN) architecture for MOS using LiDAR point-cloud sequence range images.
• The network features significantly reduced parameters (2.3 M, 66% less than state-of-the-art).
• Implemented and tested the CNN on both GPU (RTX 3090) and FPGA platforms with an NVDLA-like hardware architecture.

Main Results:

• Achieved a processing time of 35.82 ms per frame on an RTX 3090 GPU.
• Obtained an intersection-over-union (IoU) score of 51.3% on the SemanticKITTI dataset.
• Demonstrated real-time performance of 32 frames per second (fps) on an FPGA, meeting autonomous vehicle requirements.

Conclusions:

• The proposed lightweight CNN offers a significant reduction in parameters and computational load for 3D LiDAR MOS.
• The network achieves competitive accuracy while enabling real-time processing on embedded platforms.
• This advancement facilitates the deployment of robust moving object segmentation in autonomous driving systems.