C2L3-Fusion: An Integrated 3D Object Detection Method for Autonomous Vehicles

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Accurate 3D object detection is essential for the safe operation of autonomous vehicles (AVs) in complex environments.
• Existing fusion methods often face challenges with feature misalignment, impacting detection accuracy.

Purpose of the Study:

• To introduce C2L3-Fusion, a novel framework integrating YOLOv8 (2D camera) and PointPillars (3D LiDAR) for improved 3D object detection.
• To enhance spatial consistency and multi-level feature aggregation for superior detection performance.

Main Methods:

• Developed C2L3-Fusion, a novel framework fusing camera-based 2D object detection (YOLOv8) with LiDAR-based 3D object detection (PointPillars).
• Enhanced feature aggregation and spatial consistency to overcome limitations of conventional fusion techniques.
• Implemented and tested the framework on the KITTI dataset and an Nvidia Jetson AGX Xavier embedded platform.

Main Results:

• Achieved state-of-the-art performance on the KITTI dataset with mean Average Precision (mAP) scores of 89.91% (easy), 79.26% (moderate), and 78.01% (hard).
• Demonstrated superior performance compared to standalone YOLOv8, standalone PointPillars, and YoPi-CLOCs Fusion Network.
• Maintained real-time performance on embedded hardware, showcasing robustness for practical AV applications.

Conclusions:

• C2L3-Fusion offers a robust and accurate solution for 3D object detection in autonomous navigation.
• The framework's ability to enhance spatial consistency and aggregate multi-level features leads to significant improvements in detection accuracy.
• The successful real-time implementation on embedded platforms makes C2L3-Fusion highly suitable for self-driving vehicles.