Context-Aware Query Selection for Active Learning in Event Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Activity recognition is crucial for surveillance but current methods struggle with continuous learning and unlabeled data.
• Existing approaches often require fully labeled datasets and are not adaptable to evolving data streams.

Purpose of the Study:

• To develop a continuous-learning framework for context-aware activity recognition from unlabeled video.
• To reduce manual annotation effort through an intelligent active-learning technique.
• To enable online adaptation of recognition models.

Main Methods:

• A novel active-learning technique leveraging both individual activity informativeness and contextual information for query selection.
• Formulation of a conditional random field model to encode contextual relationships.
• An information-theoretic approach using entropy and mutual information for selecting informative queries.
• Graphical inference techniques for incremental model updates with human-provided labels.

Main Results:

• The proposed framework achieves superior performance in activity recognition compared to existing methods.
• Significant reduction in manual labeling effort was demonstrated.
• The framework successfully adapts learned models online with new data.

Conclusions:

• The developed continuous-learning framework offers an effective solution for context-aware activity recognition in surveillance.
• The novel active-learning strategy efficiently minimizes annotation costs while maximizing model performance.
• The system's ability for online adaptation makes it suitable for dynamic, real-world applications.