A Small-Object-Detection Algorithm Based on LiDAR Point-Cloud Clustering for Autonomous Vehicles.

Area of Science:

• Computer Vision
• Robotics
• Autonomous Systems

Background:

• 3D object detection using LiDAR point clouds is crucial for autonomous vehicle navigation.
• Existing methods struggle with accurately detecting small objects such as pedestrians and cyclists.
• Limitations in current point-cloud-based object detection hinder the safety and reliability of driverless vehicles.

Purpose of the Study:

• To develop an effective small-object detection algorithm for LiDAR point clouds.
• To enhance the detection capabilities for vulnerable road users like pedestrians and cyclists.
• To improve the overall performance and accuracy of 3D object detection systems in autonomous driving.

Main Methods:

• Proposed a novel segmented ground-point cloud segmentation algorithm utilizing multi-region plane-fitting and heuristic rules.
• Implemented an improved Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm for small object point cloud clustering.
• Incorporated K-means++ for pre-clustering and adaptively adjusted neighborhood radius, alongside an improved core point search method.
• Utilized a directional wraparound box model for final small object detection.

Main Results:

• The proposed ground segmentation algorithm achieved 91.86% precision and 92.70% recall.
• The improved DBSCAN clustering algorithm enhanced pedestrian recall by 15.89% and cyclist recall by 9.50%.
• Visualization experiments confirmed the algorithm's capability for accurate small object detection.

Conclusions:

• The developed clustering-based algorithm significantly improves the detection of small objects in LiDAR point clouds.
• The novel ground segmentation and improved DBSCAN methods offer a robust solution for autonomous vehicle perception.
• This approach enhances the safety and reliability of driverless vehicles by enabling accurate detection of pedestrians and cyclists.