3D Object Detection for Self-Driving Cars Using Video and LiDAR: An Ablation Study

Area of Science:

• Computer Vision and Robotics
• Sensor Fusion for Autonomous Systems

Background:

• High-precision 3D object detection often relies on expensive 64-beam LiDAR sensors (approx. $75,000).
• Previous research introduced SLS-Fusion, a cost-effective method fusing low-cost four-beam LiDAR with stereo cameras, outperforming advanced stereo-LiDAR fusion techniques.

Purpose of the Study:

• To analyze the contribution of stereo and LiDAR sensors to the SLS-Fusion model's 3D object detection performance based on the number of LiDAR beams.
• To quantify the stereo camera's role and its variation with respect to the number of LiDAR beams used in the fusion model.

Main Methods:

• Proposed dividing the SLS-Fusion network into two independent decoder networks to evaluate the distinct contributions of LiDAR and stereo camera components.
• Conducted experiments to assess the impact of varying LiDAR beam counts on the overall fusion model performance.

Main Results:

• The stereo camera data significantly contributes to the SLS-Fusion model's performance.
• Increasing the number of LiDAR beams beyond four showed no significant impact on the SLS-Fusion model's 3D object detection performance.

Conclusions:

• The findings suggest that cost-effective, lower-beam LiDAR sensors, when fused with stereo cameras via SLS-Fusion, can achieve high performance in 3D object detection.
• Results provide practical guidance for designing cost-efficient 3D object detection systems, indicating that expensive, high-beam LiDAR may not be necessary.