Detection of Anomalous Behavior in Prisoner Gatherings Using Surveillance Cameras

With Emphasis on Artificial Intelligence Applications

• Hadi Sadoghi Yazdi : PhD in Electronics, Expert Consultant in Machine Vision/Learning and Data Systems
• Affilation and institute:
  • Professor of Electrical and Computer Engineering, Ferdowsi University of Mashhad
  • Director of Pattern Recognition Laboratory
  • Member of SCIIP - Center of Excellence on Soft Computing and Intelligent Information Processing

Identification of Behavior in Prison with AI

Table of Contents

• Introduction
  • Objectives
  • Existing Implementations in Various Countries
• Methodology
  • Data Collection
  • Model Development
  • Deployment
  • Technology Stack
• Challenges and Mitigations
• Expected Outcomes
• Additional Service Offerings
• References
• Contact Me

Introduction

Objectives

• Develop AI-powered system for real-time anomaly detection in prison CCTV feeds.
• Reduce response times to threats (e.g., fights, self-harm).
• Ensure ethical compliance (privacy, bias mitigation).
• Achieve 90%+ accuracy in detecting crowd-based anomalies.

AI-based real-time anomaly detection in prison CCTV

Existing Implementations in Various Countries

• China: AI surveillance with facial recognition for threat detection (prison breaks impossible).

Stephen Chenin Beijing, 1 Apr 2019

• UK: Avigilon (software for contraband and behavior monitoring). Avigilon’s self-learning AI continuously improves detection accuracy by adapting to environments, reducing false alarms and enhancing security threat identification.
• Singapore: AI-based CCTV for fights and headcount checks.
• India: Facial recognition in Tihar jail for anomaly detection.

Singapore: AI-based CCTV for fights and headcount checks. New technology on trial at Changi Prison can detect cell fights through video analytics.

Fights and headcount checks

India: Tihar Jail, one of the largest prison complexes in India, has been implementing a facial recognition, anomaly system and enhance security.

Tihar installs 1,248 CCTVs with facial recognition

Methodology

Data Collection

• Gather anonymized CCTV footage (ethical approvals).
• Label data for normal vs. anomalous behaviors (e.g., fights, gatherings).
• Example: Deep Learning for Video Anomaly Detection: A Review

Illustrations of frame-level (Top) and pixel-level (Bottom) output

Model Development

• Frame Analysis
  • Use CNNs for video frame analysis, RNNs for temporal tracking.
• Frame Encoding
  • EfficientNet-B7 (pretrained on Kinetics-700) for high-resolution feature extraction.
  • 3D Convolutional Blocks (I3D) for short-term spatiotemporal features (5-frame snippets).
• Temporal Context
  • Bidirectional Quasi-Recurrent Neural Networks (QRNNs): Lightweight alternative to LSTMs/GRUs. Processes sequences with parallel convolution + recurrent pooling (reducing latency 40% vs. GRU).
  • Attention Mechanisms: Self-attention layers to weight critical frames (e.g., sudden motion/occlusion).
• Implement anomaly detection
  • use YOLO v7, Autoencoders
  • use Adaptive frame skipping based on optical flow magnitude.
• Edge Computing Deployment
  • Hardware: NVIDIA Jetson AGX Orin (48 TOPS) / Google Coral TPU.
  • Model Distillation: Teacher (EfficientNet-B7 + QRNN) → Student (MobileNetV3 + QRNN).

Deep Neural Network

Deployment

• Deploy on CCTV with AI accelerators (e.g., NVIDIA Jetson).
• Pilot testing in controlled prison settings.
• Conduct ethical audits (GDPR (General Data Protection Regulation)-like policies, bias checks).

Deployment of AI-enabled CCTV system on-site with automotive-grade accelerators for real-time monitoring.

Technology Stack

Component	Description
Hardware	High-res CCTV with IR, motion sensors
Software	Python (OpenCV, TensorFlow), MQTT alerts
AI Models	CNN-RNN, YOLO, GANs for synthetic data
Security	Encrypted storage, access controls

Technology Stack Visualization

Challenges and Mitigations

• Privacy Concerns: GDPR-like policies, data anonymization.
• AI Bias: Diverse training data, fairness algorithms.
• False Positives: Fine-tune models with prison-specific data.

Privacy, fairness, and accuracy: GDPR-like privacy with anonymization, bias mitigation through diverse data and fairness algorithms, and reduced false positives via prison-specific fine-tuning.

Expected Outcomes

• 30-50% reduction in security incidents (inspired by India).
• Cost savings via automated monitoring (like Singapore).
• Scalable system for global adoption.

Significant security gains, cost savings, and scalable global deployment

Additional Service Offerings

• Smart Prisons: Closing the gap between smart-city innovations and their application in correctional facilities.
• Tackle Prison Contraband: Trial of new technology to tackle prison contraband.
• Guarding the Perimeter: Drone Detection and Neutralization in Prisons

Protect Your Correctional Facility From Drone Threats

References

• AI surveillance systems in China, Hong Kong, Singapore (2023 reports).
• UK Altcourse prison: Avigilon software (2022).
• India Punjab jails: AI CCTV deployment (2024).

Contact Me

Hadi Sadoghi Yazdi

AI and Data Specialist

Email: h-sadoghi@um.ac.ir

Phone: +98-51-38805117

Website: https://h-sadoghi.github.io/ | https://hadisadoghiyazdi1971.github.io/