Biview learning for human posture segmentation from 3D points cloud.

Area of Science:

• Computer Vision
• Human Motion Analysis
• Machine Learning

Background:

• Posture segmentation is crucial for human motion analysis.
• Current methods use high-dimensional features, leading to memory and scalability issues.
• Efficient feature extraction is needed for large datasets.

Purpose of the Study:

• To propose an efficient two-stage dimension reduction scheme called biview learning.
• To encode depth-difference features (DDF) and relative position features (RPF) for posture segmentation.
• To create a compact yet comprehensive low-dimensional feature space.

Main Methods:

• Biview learning utilizes two stages: Discriminative Locality Alignment (DLA) for DDF and Canonical Correlation Analysis (CCA) for DDF and RPF.
• DLA reduces DDF dimensionality, while CCA explores complementary properties of DDF and RPF.
• A Support Vector Machine (SVM) classifier is trained on the CCA output.

Main Results:

• The proposed biview learning scheme effectively reduces feature dimensionality.
• The method demonstrates superior performance compared to state-of-the-art techniques.
• Validation on a 3D human point cloud dataset confirms effectiveness.

Conclusions:

• Biview learning offers an efficient and effective solution for human posture segmentation.
• The combination of DLA and CCA successfully leverages complementary features.
• This approach addresses the limitations of high-dimensional feature extraction in motion analysis.