Point Cloud Compression: Impact on Object Detection in Outdoor Contexts

Area of Science:

• Computer Vision
• Robotics
• Autonomous Systems

Background:

• Autonomous driving systems require advanced perception, including object detection, generating massive sensor data.
• Onboard processing limitations necessitate offloading computation and transmitting sensor data via vehicle-to-infrastructure (V2I) or vehicle-to-vehicle (V2V) links.
• Light Detection and Ranging (LiDAR) point clouds are crucial but voluminous, posing transmission challenges without compression.

Purpose of the Study:

• To evaluate the impact of point cloud compression on object detection performance in autonomous driving.
• To assess the influence of compression levels on object detection accuracy using different architectures.
• To analyze the effect of compression on depth maps generated from point clouds.

Main Methods:

• Object detection was performed on raw and compressed LiDAR point clouds from the KITTI dataset.
• Two distinct object detection architectures were employed.
• The study analyzed compression's impact on depth maps generated via two projection methods.

Main Results:

• Low-to-medium levels of point cloud compression showed minimal degradation in object detection performance.
• Larger objects were less affected by compression artifacts.
• Object detection using depth maps derived from compressed point clouds demonstrated competitive results against raw point cloud data.

Conclusions:

• Point cloud compression is a viable strategy for enabling efficient data transmission in autonomous driving systems.
• Depth map generation from compressed point clouds presents a robust alternative for perception tasks.
• Further research can optimize compression techniques for enhanced object detection reliability.