ExistenceMap-PointPillars: A Multifusion Network for Robust 3D Object Detection with Object Existence Probability...

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Object recognition is vital for automated driving safety.
• LiDAR-camera fusion offers richer data for 3D object detection than single sensors.
• Existing methods struggle with stable recognition in adverse conditions like night or rain.

Purpose of the Study:

• To introduce ExistenceMap-PointPillars, a novel approach for robust 3D object detection.
• To improve 3D object detection performance, particularly under challenging environmental conditions.
• To enhance the reliability of automated driving systems through advanced sensor fusion.

Main Methods:

• Modified a LiDAR-based 3D object detection network (PointPillars).
• Integrated pseudo 2D maps estimating object existence regions from fused sensor data.
• Incorporated these maps into a pseudo image generated from 3D point clouds.

Main Results:

• Achieved a +4.19% improvement in mean Average Precision (mAP) over conventional PointPillars.
• Demonstrated enhanced focus on object existence regions using Grad-CAM analysis.
• Showcased a reduction in false positives, indicating improved detection accuracy.

Conclusions:

• ExistenceMap-PointPillars offers a significant advancement in 3D object detection.
• The method provides improved performance and robustness, especially in adverse conditions.
• This approach contributes to safer and more reliable automated driving systems.