Supervised preserving projection for learning scene information based on time-of-flight imaging sensor

Area of Science:

• Computer Vision
• Machine Learning
• 3D Imaging

Background:

• Time-of-flight (TOF) sensors provide depth images but suffer from nonstatic noise and distance ambiguity.
• Existing manifold learning techniques may not fully capture the complex structures in TOF data.
• Local surface patches offer a robust representation for approximating manifold structures in noisy TOF data.

Purpose of the Study:

• To propose a novel supervised manifold learning approach, Supervised Preserving Projection (SPP), for TOF depth images.
• To enhance the learning of scene information from TOF data by leveraging local surface patch structures.
• To develop a method robust to nonstatic noise inherent in TOF 3D imaging sensors.

Main Methods:

• Utilizing local surface patches to approximate underlying manifold structures, offering robustness against TOF data noise.
• Implementing SPP to preserve pairwise similarity between local neighboring patches in TOF depth images.
• Achieving low-dimensional embedding by incorporating scene region class labels and local geometrical properties.

Main Results:

• SPP effectively learns scene information from TOF depth images across various scenarios.
• The proposed method demonstrates robustness to nonstatic noise present in 3D imaging sensor data.
• Real-time estimation results are achieved through low-dimensional embedding and predictive mapping.

Conclusions:

• SPP provides an effective supervised manifold learning approach for TOF depth image analysis.
• The method successfully addresses challenges associated with TOF data, such as noise and ambiguity.
• SPP offers advantages over classical linear and nonlinear manifold learning techniques for 3D imaging applications.