Prediction of Human Activity by Discovering Temporal Sequence Patterns.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Early prediction of human activities is crucial for time-critical applications.
• Complex human activities involve intricate temporal compositions of simple actions and object interactions.

Purpose of the Study:

• To propose a novel framework for long-duration complex activity prediction.
• To address the challenge of predicting activities by discovering causality, context-cues, and predictability.

Main Methods:

• Developed a general framework for complex activity prediction using temporal sequence pattern mining.
• Employed probabilistic suffix trees (PST) to model causal relationships and Markov dependencies between actions.
• Utilized sequential pattern mining (SPM) to encode context-cues, including object interactions, as symbolic sequences.
• Introduced a predictive accumulative function (PAF) to quantify activity predictability.

Main Results:

• The framework demonstrated superior performance in predicting global activity classes and local action units.
• Evaluated effectiveness across two scenarios: action-only prediction and context-aware prediction, using two datasets for each.
• Successfully captured both large and small order Markov dependencies in action sequences.

Conclusions:

• The proposed framework offers a systematic approach to complex activity prediction.
• Integrating causality, context-cues (especially object interactions), and predictability enhances prediction accuracy.
• The method shows significant potential for real-world applications requiring accurate human activity recognition.