Event-Based 3D Motion Flow Estimation Using 4D Spatio Temporal Subspaces Properties

Area of Science:

• Computer Vision
• Robotics
• Sensor Fusion

Background:

• Current scene flow estimation relies on image luminance, limiting its application.
• Neuromorphic sensors offer asynchronous, continuous data streams.

Purpose of the Study:

• To develop a purely time-based scene flow estimation method for 3D point clouds.
• To create a unifying framework for synchronous and asynchronous 3D data.
• To enable scene flow estimation without luminance information.

Main Methods:

• Formulating scene flow estimation using local piecewise regularization.
• Utilizing 4D space-time properties and subspace decomposition.
• Exploiting asynchronous event-based vision sensor data for continuous reconstruction.

Main Results:

• Successful scene flow estimation from synchronous and asynchronous 3D point clouds.
• Demonstrated capability to handle deformable object motion.
• Validated performance across diverse 3D sensors.

Conclusions:

• The proposed time-based method offers a robust alternative to luminance-dependent techniques.
• The framework effectively integrates data from various 3D sensors, including neuromorphic ones.
• This approach advances scene flow estimation for dynamic and complex environments.