An Online Continuous Human Action Recognition Algorithm Based on the Kinect Sensor.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Continuous human action recognition (CHAR) is crucial for natural human-robot interaction.
• Existing CHAR methods often require predefined action start and end points, limiting practicality.
• Skeletal data from RGB-D sensors offers a robust representation for human motion analysis.

Purpose of the Study:

• To propose an online CHAR algorithm for practical human-robot interactions.
• To develop a method that recognizes continuous actions without prior knowledge of their boundaries.
• To enhance the efficiency and effectiveness of CHAR using skeletal data.

Main Methods:

• Utilizing skeletal data extracted from RGB-D images via Kinect sensors.
• Modeling human actions as ordered sequences of key poses and atomic motions.
• Employing an online segmentation method based on feature potential differences to identify pose and motion segments.
• Implementing an online classification using a variable-length Maximal Entropy Markov Model (MEMM) for action recognition.

Main Results:

• The proposed algorithm effectively segments feature sequences into pose and motion components.
• Online model matching computes likelihood probabilities for key poses and atomic motions.
• The variable-length MEMM achieves accurate and efficient recognition of continuous human actions.
• Experimental results on public datasets validate the algorithm's effectiveness and high efficiency.

Conclusions:

• The developed online CHAR algorithm is effective and efficient for real-world applications.
• The method's ability to recognize actions without start/end point detection significantly advances CHAR.
• This approach holds promise for improving the naturalness and intuitiveness of human-robot interactions.