Hierarchical Relaxed Partitioning System for Activity Recognition.

Area of Science:

• Computer Vision
• Human Activity Recognition
• Machine Learning

Background:

• Recognizing human activities from motion data is challenging due to overlapping features in the analysis space.
• Existing methods may struggle with distinguishing similar activities or handling complex motion patterns.

Purpose of the Study:

• To propose a novel hierarchical relaxed partitioning system (HRPS) for robust human activity recognition.
• To address the challenge of multiple overlaps in feature spaces during activity classification.

Main Methods:

• Developed two feature descriptors from 2-D stick figure human motion analysis for cyclic and noncyclic activities.
• Implemented a hierarchical system to first classify activities as 'pure' (no overlaps) or 'impure' (multiple overlaps).
• Utilized a majority voting scheme to resolve overlaps in impure activities.

Main Results:

• The HRPS demonstrated robust recognition of various activities across low and high-resolution datasets with different views.
• The system effectively handles feature spaces with multiple overlaps, a common challenge in activity recognition.
• Validation on diverse datasets confirmed the method's accuracy and reliability.

Conclusions:

• The proposed HRPS provides an effective solution for recognizing similar human activities with overlapping features.
• HRPS offers advantages in real-time speed, ease of implementation, extensibility, and minimal training requirements.
• This approach advances human motion analysis and activity recognition systems.