PolarFusion: A multi-modal fusion algorithm for 3D object detection based on polar coordinates

Area of Science:

• Computer Vision and Robotics
• Autonomous Systems
• Machine Learning for Perception

Background:

• Current 3D object detection methods using multi-modal fusion often rely on Cartesian coordinates.
• This coordinate system can result in asymmetrical feature representation and imbalanced attention across sensor views.
• These limitations hinder the accuracy and efficiency of object detection, particularly for autonomous driving applications.

Purpose of the Study:

• To introduce PolarFusion, the first multi-modal fusion algorithm for Bird's-Eye View (BEV) object detection utilizing polar coordinates.
• To enhance the efficiency and accuracy of 3D object detection by addressing feature misalignment and improving information fusion.
• To advance the environmental perception capabilities of autonomous vehicles.

Main Methods:

• Developed PolarFusion, a novel BEV object detection algorithm operating in polar coordinates.
• Implemented three key modules: Polar Region Candidates Generation, Polar Region Query Generation, and Polar Region Information Fusion.
• Utilized region proposal-based segmentation for efficient image processing and integrated segmented regions with point cloud data to mitigate feature misalignment. Employed self-attention for effective fusion of image and point cloud data.

Main Results:

• PolarFusion achieved a NuScenes Detection Score (NDS) of 76.1% and a mean Average Precision (mAP) of 74.5% on the nuScenes test set.
• Demonstrated significant performance improvements over existing Cartesian-based multi-modal fusion methods.
• Qualitative and quantitative results confirmed the algorithm's effectiveness in challenging BEV object detection scenarios.

Conclusions:

• PolarFusion represents a significant advancement in multi-modal 3D object detection by leveraging polar coordinates.
• The proposed approach effectively fuses sensor information, leading to superior performance in BEV object detection.
• This work contributes to enhanced environmental perception for autonomous vehicles and the development of intelligent transportation systems.