Rejection of Irrelevant Human Actions in Real-time Hidden Markov Model based Recognition Systems for Wearable...

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Hidden Markov Models (HMMs) are effective for pattern detection in observation sequences.
• HMMs struggle with unseen patterns, leading to false positives in unconstrained environments.
• Existing methods may misclassify novel patterns as similar to trained ones.

Purpose of the Study:

• To develop a threshold-based technique for early detection and classification of unwanted patterns in continuous observation streams.
• To reduce false positives and improve the robustness of HMMs in dynamic environments.
• To enhance human action recognition accuracy in body sensor networks.

Main Methods:

• A threshold-based validation approach is proposed for detecting irrelevant patterns.
• Test patterns are validated against fixed-length substrings of observation sequences.
• Similarity comparison using a threshold value determines pattern validity.

Main Results:

• The technique achieves approximately 93% accuracy in detecting unwanted actions.
• It maintains a high accuracy of 94% in recognizing valid human actions.
• Early detection of unwanted patterns is enabled, reducing latency.

Conclusions:

• The proposed threshold-based method effectively identifies and classifies unwanted patterns in observation streams.
• This approach enhances HMM performance by mitigating false positives and improving detection speed.
• The technique shows significant promise for applications like body sensor network-based human action recognition.