Attention-based bidirectional-long short-term memory for abnormal human activity detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Abnormal human behavior poses societal risks, including assault and online hate crimes.
• Existing methods for monitoring and controlling aberrant activities are often impractical or unworkable.
• Deep learning's potential for categorizing anomalous human activity from video data is largely unexplored.

Purpose of the Study:

• To develop and evaluate a deep learning framework for detecting and categorizing abnormal human behavior in video streams.
• To address the limitations of current approaches in monitoring and controlling societal risks associated with aberrant activities.

Main Methods:

• A novel deep learning architecture combining a convolutional neural network (CNN), bidirectional long short-term memory (Bi-LSTM), and an attention mechanism was implemented.
• The framework analyzes spatiotemporal characteristics of raw video streams to detect anomalous human activity.
• The model is designed to reliably assign detected abnormal behavior to its specific category.

Main Results:

• The proposed framework achieved high detection accuracy across multiple datasets.
• Specific accuracies include 98.9% on UCF11, 96.04% on UCF50, and 61.04% on the subUCF crime dataset.
• Performance was compared against state-of-the-art algorithms, demonstrating the effectiveness of the proposed architecture.

Conclusions:

• The developed deep learning approach demonstrates significant potential for accurately detecting and categorizing abnormal human behavior in videos.
• This framework offers a practical and workable solution for monitoring and mitigating societal harm caused by aberrant activities.
• Further research can explore the application of this model in real-world scenarios for enhanced public safety and security.