Video-based human activity recognition using multilevel wavelet decomposition and stepwise linear discriminant...

Area of Science:

• Computer Vision
• Pattern Recognition
• Signal Processing

Background:

• Human Activity Recognition (HAR) from video is a critical area in computer vision.
• Existing methods face challenges in accurately analyzing complex human motions and behaviors.
• Developing robust and accurate HAR systems remains an exigent research goal.

Purpose of the Study:

• To present a novel and accurate video-based human activity recognition system named WS-HAR.
• To leverage wavelet transform, stepwise linear discriminant analysis (SWLDA), and hidden Markov models (HMM) for enhanced HAR.
• To demonstrate the effectiveness and superiority of the proposed WS-HAR system.

Main Methods:

• Feature extraction using Symlet wavelet transform from video activity frames.
• Feature selection via stepwise linear discriminant analysis (SWLDA) based on regression values.
• Activity classification using the Hidden Markov Model (HMM) sequential classifier.

Main Results:

• The WS-HAR system achieved a weighted average recognition rate of 97% across two standard datasets.
• The proposed SWLDA technique effectively selected localized and discriminating features.
• Performance was validated using n-fold cross-validation and comparative experiments.

Conclusions:

• WS-HAR offers a robust and accurate solution for video-based human activity recognition.
• The combination of wavelet transform, SWLDA, and HMM significantly improves classification accuracy.
• The system demonstrates a notable advancement over existing statistical and state-of-the-art HAR methods.