3D-DIoU: 3D Distance Intersection over Union for Multi-Object Tracking in Point Cloud.

Area of Science:

• Computer Vision
• Point Cloud Processing
• Multi-Object Tracking (MOT)

Background:

• Multi-object tracking (MOT) in point clouds faces challenges with occlusion and similar objects, degrading performance.
• Current 'tracking-by-detection' methods suffer from increased ID switches and tracking drifts in complex scenes.
• Predicting object locations accurately in occluded or complex scenarios remains difficult.

Purpose of the Study:

• To enhance the robustness and speed of 3D multi-object tracking.
• To address performance degradation caused by occlusion and object similarity in point cloud data.
• To improve the accuracy of object detection and association in challenging tracking scenarios.

Main Methods:

• Employed a novel 3D distance IoU non-maximum suppression (DIoU-NMS) for accurate 3D object detection.
• Utilized 3D-DIoU for robust object association, improving tracking reliability.
• Developed a hybrid approach combining 3D DIoU-NMS and 3D-DIoU for efficient tracking.

Main Results:

• Significant improvement in tracking speed was achieved through the hybrid 3D DIoU-NMS and 3D-DIoU method.
• Experimental results on Waymo Open Dataset and nuScenes dataset show clear benefits over existing algorithms.
• The proposed multistage data association and tracking technique demonstrated superior tracking accuracy.

Conclusions:

• The developed 3D MOT method significantly enhances tracking accuracy and speed.
• The hybrid DIoU-NMS and 3D-DIoU approach effectively tackles challenges like occlusion and object similarity.
• This technique offers a substantial advancement compared to current 3D MOT methods.