3D-DFM: Anchor-Free Multimodal 3-D Object Detection With Dynamic Fusion Module for Autonomous Driving

Area of Science:

• Computer Vision
• Robotics
• Sensor Fusion

Background:

• Anchor-based methods dominate cross-modal 3D object detection.
• Challenges include complex parameter tuning and slow postprocessing, hindering real-time applications like autonomous driving.

Purpose of the Study:

• To develop an efficient, anchor-free architecture for camera-LiDAR 3D object detection.
• To improve the performance of embedded systems in autonomous driving.

Main Methods:

• Proposed a dynamic fusion module (DFM) for adaptive interaction between image and LiDAR point features.
• Introduced 3D distance intersection-over-union (3D-DIoU) loss for optimized 3D bounding box regression.
• Integrated components into an end-to-end multimodal 3D detector (3D-DFM).

Main Results:

• The 3D-DFM architecture demonstrated superior and universal performance on the KITTI dataset.
• Achieved competitive detection accuracy with real-time inference speeds.

Conclusions:

• This work presents the first anchor-free pipeline for multimodal 3D object detection.
• The proposed 3D-DFM offers an efficient and accurate solution for real-time 3D object detection in embedded systems.