3D Object Detection Based on Attention and Multi-Scale Feature Fusion.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Accurate 3D object detection from point clouds is crucial for safe autonomous driving.
• Existing methods often struggle with complex scenes and diverse object types.

Purpose of the Study:

• To propose MA-MFFC, a novel method for enhancing 3D object detection accuracy in point clouds.
• To improve the robustness and feature extraction capabilities of 3D object detection networks.

Main Methods:

• The proposed MA-MFFC method integrates a multi-attention (MA) module with point-channel and voxel attention mechanisms.
• A multi-scale feature fusion network (MFFC) incorporating the ConvNeXt module is employed for enhanced feature extraction.
• Attention mechanisms are applied during voxelization and within the 3D backbone to refine point and voxel features.

Main Results:

• The MA-MFFC method achieved an average accuracy of 64.60% for pedestrian detection and 80.92% for cyclist detection on the KITTI dataset.
• This represents a significant improvement of 1.33% for pedestrians and 2.1% for cyclists compared to the baseline network.
• The method demonstrated enhanced detection and localization of challenging objects.

Conclusions:

• The MA-MFFC method effectively improves 3D object detection accuracy and robustness in point cloud data.
• The integration of attention mechanisms and ConvNeXt modules contributes to superior feature representation.
• This advancement holds promise for enhancing the safety and reliability of autonomous driving systems.