Sparse Representations for Object- and Ego-Motion Estimations in Dynamic Scenes.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Autonomous systems require disentangling object and ego-motion for navigation.
• Existing methods fail to separate pixel velocities into distinct motion components.
• Unconstrained scene depth poses challenges for traditional ego-motion estimation.

Purpose of the Study:

• To develop a learning-based approach for predicting ego-motion parameters and object-motion fields (OMF).
• To achieve robustness to unconstrained scene depth variations.
• To improve upon existing methods for pixelwise motion estimation.

Main Methods:

• Utilized a convolutional autoencoder for motion prediction.
• Implemented continuous ego-motion constraints for depth-independent parameter solving.
• Introduced a novel differentiable sparsity penalty for learning an ego-motion field (EMF) basis set.
• Developed a method to extract OMF by comparing predicted EMF with optic flow.

Main Results:

• The proposed model effectively separates ego-motion and object-motion components.
• Demonstrated robustness to unconstrained scene depth.
• Achieved favorable performance on pixelwise object- and ego-motion estimation tasks.
• Outperformed state-of-the-art baselines on real and synthetic datasets.

Conclusions:

• The novel approach successfully disentangles motion sources in dynamic scenes.
• The method provides accurate ego-motion estimation and OMF extraction.
• This work advances capabilities for autonomous navigation and object tracking.