Continuous human action recognition using depth-MHI-HOG and a spotter model

Area of Science:

• Computer Vision
• Human Action Recognition
• Machine Learning

Background:

• Continuous human action recognition is challenging due to complex motion patterns.
• Existing methods often struggle with precise action segmentation and noise filtering.

Purpose of the Study:

• To propose a robust method for spotting and recognizing continuous human actions.
• To enhance the accuracy of action recognition by precisely identifying action start and end points.

Main Methods:

• A novel Depth-MHI-HOG (DMH) feature extraction method is proposed for foreground segmentation.
• Action modeling uses k-means clustering for action sequence generation, feeding into Hidden Markov Models (HMMs).
• An action spotting model filters irrelevant actions and determines precise action boundaries.

Main Results:

• The proposed method effectively separates foreground from background using depth information.
• Action spotting significantly improves the performance of continuous human action recognition.
• Experimental results demonstrate the method's efficiency in real-world environments.

Conclusions:

• The integrated approach of DMH, action modeling, and spotting enhances continuous human action recognition.
• The method provides accurate segmentation and recognition of actions in dynamic settings.
• This work offers an efficient solution for real-time human action analysis.