Efficient three-dimensional point cloud object detection based on improved Complex-YOLO.

Area of Science:

• Computer Vision
• Robotics
• Autonomous Systems

Background:

• Real-time 3D object detection and classification using LiDAR is crucial for autonomous driving.
• Existing methods like Complex-YOLO face challenges with sparse 3D data, lacking height detection and struggling with small objects.
• The sparsity and disorder of point cloud data present significant hurdles for accurate inference.

Purpose of the Study:

• To improve the accuracy and capabilities of LiDAR-based 3D object detection for autonomous vehicles.
• To address the limitations of Complex-YOLO, specifically its lack of height detection and poor performance on small objects.
• To develop a more robust and efficient algorithm for real-time 3D perception.

Main Methods:

• Enhanced Complex-YOLO by incorporating a multi-scale feature fusion network for improved small object detection.
• Replaced the backbone network with RepVGG to increase network depth and overall detection performance.
• Integrated an effective height detection module to enhance the algorithm's 3D perception capabilities.

Main Results:

• The improved algorithm demonstrated strong accuracy on the KITTI dataset.
• Achieved superior detection speeds: 48 FPS on RTX3070Ti and 20 FPS on GTX1060.
• Maintained efficient memory usage at 841 MiB.

Conclusions:

• The proposed enhancements significantly improve LiDAR-based 3D object detection accuracy, particularly for small objects and height estimation.
• The refined algorithm offers a compelling balance of high accuracy and real-time processing efficiency for autonomous driving applications.
• The method provides a robust solution for real-time 3D perception challenges posed by sparse LiDAR data.