MonOri: Orientation-Guided PnP for Monocular 3-D Object Detection

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Robotics

Background:

• Monocular 3-D object detection is crucial for autonomous driving but struggles with occluded objects.
• Existing methods often use pseudolabels, neglecting keypoint geometric relationships, leading to poor performance on occlusions.

Purpose of the Study:

• To enhance monocular 3-D object detection performance, particularly for occluded objects.
• To leverage object orientation information within the perspective-n-point (PnP) algorithm for improved spatial estimation.

Main Methods:

• Developed MonOri, an orientation-guided PnP method for monocular 3-D object detection.
• Introduced Feature Aggregation Detection Module (FADM) with Feature Focus Fusion Module (FFFM) and CondConv Detection Module (CCDM) to handle object deformation and occlusion.
• Proposed Orientation-Guided Keypoints' Selection Module (OGKSM) for accurate keypoint optimization and spatial inference.

Main Results:

• MonOri achieved competitive results in monocular 3-D object detection.
• Demonstrated that incorporating orientation information into the PnP algorithm mitigates occlusion's impact.
• Significantly improved the recognition rate of occluded objects.

Conclusions:

• Object orientation is a valuable cue for improving 3-D object detection under occlusion.
• The MonOri method offers a robust solution for detecting occluded objects in autonomous driving scenarios.
• The proposed FADM and OGKSM modules effectively address challenges posed by object deformation and occlusion.