S2*-ODM: Dual-Stage Improved PointPillar Feature-Based 3D Object Detection Method for Autonomous Driving.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Current PointPillar methods for 3D object detection struggle with under-segmentation, overlapping objects, and false detections, especially in occluded environments.
• Accurate 3D object detection is critical for the safety and reliability of autonomous driving systems.

Purpose of the Study:

• To develop a novel dual-stage improved PointPillar feature-based 3D object detection method (S2*-ODM) to overcome limitations of existing approaches.
• To enhance the recognition of local structures and global distributions for better object differentiation in complex scenarios.

Main Methods:

• Introduced a dual-stage pillar feature encoding (S2-PFE) module integrating inter-pillar and intra-pillar relational features.
• Incorporated an attention mechanism into the backbone network to refine feature extraction by emphasizing critical features and suppressing irrelevant ones.

Main Results:

• The S2*-ODM method demonstrated superior performance over the baseline on the KITTI dataset.
• Achieved notable improvements in detection accuracy: +1.04% for cars, +2.17% for pedestrians, and +3.72% for cyclists.

Conclusions:

• The proposed S2*-ODM significantly enhances the accuracy and reliability of 3D object detection for autonomous driving.
• The dual-stage pillar feature encoding and attention mechanism are key innovations for improved performance in occluded and overlapping scenarios.