Human action recognition based on kinematic similarity in real time

Area of Science:

• Computer Vision
• Pattern Recognition
• Robotics

Background:

• Human action recognition is crucial for human-robot interaction and pattern recognition.
• 3D pose data acquisition hardware is now widely available, driving interest in this field.

Purpose of the Study:

• To propose a fast, simple, and powerful human action recognition method.
• To address challenges with unsegmented and irregularly segmented datasets.

Main Methods:

• Utilizes human kinematic similarity with joint positions, angular velocity, and acceleration.
• Employs a short sliding time window (approx. 5 frames) for pose representation.
• Incorporates three modified k-nearest neighbors (KNN) classifiers for frame confidence, labeling, and action classification.

Main Results:

• The method is efficient and capable of real-time performance.
• Successfully regularizes irregularly segmented public datasets.
• Achieves higher accuracy on challenging datasets like MSR-Action3D, MSRDailyActivity3D, and UTD-MHAD.

Conclusions:

• The proposed method offers an effective solution for human action recognition from 3D pose data.
• Demonstrates robustness in handling unsegmented and difficult action sequences.
• Provides a practical approach for real-time applications and dataset regularization.