Real-Time Violence and Anomaly Detection from CCTV Video Streams using Deep Learning

To design a deep learning framework for automatic detection of violent and anomalous activities from CCTV footage.
To develop video-based models capable of learning spatial and temporal patterns of human behavior (violence).
To classify video segments into normal and abnormal activity categories.
To generate real-time alerts when suspicious or violent events are detected.
To integrate explainable AI techniques to highlight visual regions influencing model decisions.
To build a prototype dashboard for monitoring live or recorded CCTV streams.

Complexity of Human Behavior: Human actions vary widely in appearance, speed, and context, making anomaly detection difficult.
Lack of Labeled Data: Large, well-annotated datasets for violent and abnormal events are limited.
High Computational Cost: Video processing requires significant memory and processing power.
False Alarms: Normal actions such as running or crowd movement may be misclassified as violent.
Lighting and Camera Variations: CCTV footage often suffers from poor lighting, low resolution, and occlusions.
Privacy and Ethical Concerns: Surveillance data must be handled carefully to ensure anonymity and responsible usage.

Team Lead

Dr. Raj Gaurav Mishra

rajgaurav.mishra@nmims.edu

Team Members