MAT-PointPillars: Enhanced PointPillars algorithm based on multi-scale attention mechanisms and transformer

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Existing 3D object detection algorithms struggle with small, irregular targets like cyclists.
• Low detection accuracy and recognition errors hinder the reliability of autonomous systems.

Purpose of the Study:

• To develop an improved 3D object detection algorithm, MAT-PointPillars, specifically addressing the challenges of detecting small and irregular objects.
• To enhance the accuracy and robustness of 3D target detection for applications like autonomous driving.

Main Methods:

• MAT-PointPillars extends the PointPillars algorithm by incorporating multi-scale vision Transformers and attention mechanisms.
• The method utilizes pillar coding for semantic point cloud encoding and integrates attention mechanisms to refine the backbone's upsampling process.
• A Transformer Encoder is introduced to enhance the upsampling structure in the third stage of the backbone.

Main Results:

• On the KITTI dataset, MAT-PointPillars achieved 3D average detection accuracy (AP3D) of 81.15%, 62.02%, and 58.68% across three difficulty levels.
• The algorithm demonstrated improvements in AP3D by 2.44%, 1.19%, and 1.23% compared to the baseline model.
• A real-time 3D object detection system built with ROS achieved an average running speed of 22.63 frames per second.

Conclusions:

• MAT-PointPillars significantly improves the detection accuracy of small and irregular objects, particularly cyclists, in 3D.
• The algorithm maintains a high detection speed, suitable for real-time applications and exceeding conventional LiDAR sampling frequencies.
• The enhanced system offers increased reliability for autonomous systems operating in complex, real-world scenarios.