AMFF-Net: An Effective 3D Object Detector Based on Attention and Multi-Scale Feature Fusion

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Autonomous vehicle applications increasingly rely on accurate LiDAR point cloud 3D object detection.
• Voxel-based feature aggregation methods show promise but struggle with background point filtering and small object detection.
• Existing methods often lack efficiency and precision in complex 3D scenes.

Purpose of the Study:

• To propose an Attention-based and Multiscale Feature Fusion Network (AMFF-Net) for enhanced LiDAR 3D object detection.
• To improve the detection of small objects and reduce computational overhead in autonomous driving.
• To enhance the precision and efficiency of 3D object detection in complex environments.

Main Methods:

• Developed a Dual-Attention Voxel Feature Extractor (DA-VFE) incorporating pointwise and channelwise attention to refine voxel features.
• Implemented a Multi-scale Feature Fusion (MFF) Module with self-calibrated convolutions, residual structure, and coordinate attention as a 2D backbone.
• Integrated DA-VFE and MFF Module to enhance feature extraction and contextual information capture.

Main Results:

• AMFF-Net achieved 62.8% mAP on the nuScenes dataset, significantly improving small object detection performance.
• The proposed network demonstrated a significant reduction in computational overhead compared to baseline methods.
• Maintained comparable inference speed while enhancing detection accuracy.

Conclusions:

• AMFF-Net effectively addresses limitations in background point filtering and small object detection for LiDAR 3D object detection.
• The network offers a superior balance of precision, efficiency, and computational cost for autonomous driving systems.
• Achieved state-of-the-art performance on both nuScenes and KITTI datasets.