3D Object Detection under Urban Road Traffic Scenarios Based on Dual-Layer Voxel Features Fusion Augmentation

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Deep Learning Architectures

Background:

• Accurate object detection is crucial for intelligent vehicles in complex urban environments.
• Challenges include object misrecognition due to occlusion and limited fields of view.

Purpose of the Study:

• To propose a novel network, the dual-layer voxel feature fusion augmentation network (DL-VFFA), for enhanced object detection.
• To address limitations in current methods regarding occlusion and field-of-view constraints.

Main Methods:

• Employs a point cloud voxelization architecture with Mahalanobis distance for point cloud association.
• Integrates local and global information via weight sharing and computes relative position encoding using an attention Gaussian deviation matrix.
• Features a two-layer feature fusion mechanism (voxel-to-voxel and point cloud-to-image) with learnable weights.

Main Results:

• DL-VFFA demonstrates significant performance improvements on the KITTI dataset compared to the baseline Second network.
• Outperforms baseline in medium- and high-difficulty scenarios, excelling in capturing fine-grained object features post-voxelization.
• Ablative studies confirm the effectiveness of the proposed voxel fusion modules.

Conclusions:

• The DL-VFFA network offers a robust solution for enhancing object detection accuracy in intelligent vehicles.
• The proposed fusion strategies effectively handle challenges posed by occlusion and limited views, improving overall system reliability.