Muti-Frame Point Cloud Feature Fusion Based on Attention Mechanisms for 3D Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Continuous frame point-cloud object detection is an emerging research area.
• Current multi-frame point cloud fusion methods often rely on GPS/IMU for alignment, limiting their effectiveness to static objects.
• Existing methods struggle to accurately detect moving objects in dynamic environments.

Purpose of the Study:

• To develop an advanced multi-scale feature fusion method for point-cloud object detection.
• To enable accurate detection of both static and moving objects without external registration data (GPS/IMU).
• To address the computational inefficiency of non-local methods in point cloud processing.

Main Methods:

• Proposed a non-local-based multi-scale feature fusion approach for continuous point clouds.
• Introduced a simplified non-local block leveraging point cloud sparsity, reducing memory consumption by 99.93%.
• Implemented triple attention mechanisms to enhance object features and suppress background noise.

Main Results:

• The proposed method achieved improved mean Average Precision (mAP) by 3.9% and 4.1% compared to concatenation-based fusion on PointPillars-2 and CenterPoint-2, respectively.
• Outperformed the state-of-the-art 3D-VID method by 1.2% in mAP.
• Demonstrated effective handling of both static and moving objects without GPS/IMU registration.

Conclusions:

• The novel non-local feature fusion method offers a significant advancement in continuous frame point-cloud object detection.
• The simplified non-local block and triple attention effectively improve detection accuracy and computational efficiency.
• This approach provides a robust solution for real-world applications requiring accurate 3D object detection from point cloud data.