Voxel self-attention and center-point for 3D object detector

Area of Science:

• Computer Vision
• Robotics
• Autonomous Systems

Background:

• The increasing demand for autonomous driving necessitates advanced 3D object detection.
• Current convolutional neural network (CNN)-based, anchor-dependent LiDAR detectors face limitations in scene comprehension and precise object orientation prediction.

Purpose of the Study:

• To address the limitations of existing 3D object detection methods for autonomous driving.
• To improve the accuracy of object detection and posture estimation in LiDAR data.

Main Methods:

• Proposed Voxel Self-Attention and Center-point (VSAC) method.
• Incorporated a voxel self-attention network to capture comprehensive voxel relationships.
• Introduced a pseudo spatiotemporal feature pyramid net (PST-FPN) to optimize feature weights.
• Utilized a center-point detection head for more accurate steering predictions.

Main Results:

• VSAC demonstrated positive performance on benchmark datasets.
• The method effectively captures extensive voxel relationships.
• Improved prediction accuracy for object posture during steering.

Conclusions:

• VSAC offers a promising advancement in 3D object detection for autonomous driving.
• The proposed architecture overcomes limitations of traditional CNN-based and anchor-dependent methods.
• VSAC shows strong potential for real-world autonomous vehicle applications.