Dynamic Rigid Bodies Mining and Motion Estimation Based on Monocular Camera.

Area of Science:

• Robotics and Computer Vision
• Dynamic Object Perception
• Motion Estimation

Background:

• Robot navigation in dynamic environments is challenging without prior object or motion knowledge.
• Existing methods often rely on sampling and reprojection errors for motion hypothesis association.
• Distinguishing dynamic objects from static backgrounds is crucial for robust Simultaneous Localization and Mapping (SLAM).

Purpose of the Study:

• To propose a novel method for dynamic rigid body detection and motion estimation using a monocular camera.
• To improve the accuracy and robustness of motion estimation in dynamic environments.
• To enable SLAM initialization in scenarios with moving objects.

Main Methods:

• Introduced a probabilistic field on the Sim(3) manifold to represent relevance among motion hypotheses.
• Utilized random sampling to establish the probabilistic field and identify regions of potential rigid bodies.
• Developed a pose calculation method based on feature points from high-confidence regions, reducing sampling randomness.

Main Results:

• The proposed method effectively mines dynamic rigid bodies and estimates their motion from monocular sequences.
• It enhances inlier detection for rigid bodies, leading to more accurate motion estimation.
• Experimental validation on KITTI, Hopkins 155, and MTPV62 datasets demonstrated superior performance in dynamic object perception.

Conclusions:

• The novel approach accurately perceives dynamic objects and estimates their motion, outperforming existing methods.
• This technique facilitates the initialization of SLAM in complex, dynamic environments.
• The probabilistic field on the Sim(3) manifold offers a computable way to represent motion hypothesis relevance.