Spatio-temporal filter for dense real-time Scene Flow estimation of dynamic environments using a moving RGB-D camera

• Scene Flow estimation is fast while avoiding over-smoothing objects boundaries.
• Occlusion problem and sensor noise are avoided without any hardware modification.
• Adaptive spatial filter improves the quality of the device's depth and Optical Flow.
• 3D Kalman filter is used for temporal smoothness and robustness at object edges.

In this paper, we present an automated method for dense real-time Scene Flow estimation of dynamic scenes using Microsoft's depth sensor Kinect. The main contribution of the proposed method is that the estimation is fast while avoiding over-smoothing objects boundaries, occlusion problem and sensor noise without any hardware modification. In particular, the proposed method improves the quality of the device's depth and computed Optical Flow by applying an adaptive spatial filter combined with 3D Kalman filter for temporal smoothness and robustness at object edges. Quantitative evaluations show that the proposed method can produce Scene Flow with higher accuracy and low computational time compared to the state-of-the-art methods.