HFSA-Net: A 3D Object Detection Network with Structural Encoding and Attention Enhancement for LiDAR Point Clouds.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• LiDAR point cloud data is inherently sparse, posing challenges for 3D object detection.
• Existing feature encoding methods, particularly voxelization, struggle to preserve crucial geometric information from sparse point clouds, impacting detection performance.

Purpose of the Study:

• To propose an enhanced 3D object detection framework that effectively addresses the limitations of sparse LiDAR data.
• To improve the retention of geometric structural information during feature encoding.
• To enhance the accuracy and generalization ability of 3D object detection models.

Main Methods:

• Structured Voxel Feature Encoder for intra-voxel refinement and multi-scale context aggregation.
• Hybrid-Domain Attention-Guided Sparse Backbone with decoupled hybrid attention and hierarchical integration.
• Scale-Aggregation Head utilizing multi-level feature pyramid fusion and cross-layer interaction.

Main Results:

• The proposed framework achieved a 3.34% increase in mean Average Precision (mAP) on the KITTI dataset compared to the baseline.
• Demonstrated improved 3D detection accuracy and generalization on a vehicle platform with lower-resolution LiDAR.

Conclusions:

• The enhanced framework effectively improves 3D object detection from sparse LiDAR data.
• The proposed methods for feature encoding, attention guidance, and scale aggregation contribute to superior performance and robustness.