Deep Reinforcement Learning-Empowered Cost-Effective Federated Video Surveillance Management Framework.

Area of Science:

• Computer Vision and Machine Learning
• Edge Computing Architectures
• Artificial Intelligence for Security

Background:

• Deep learning (DL) enhances video surveillance precision but demands significant computational and memory resources (GPU power and memory).
• Existing methods for managing DL models in surveillance often use static thresholds or moving averages, which are inefficient.
• The need for real-time object tracking and behavior analysis in video surveillance systems creates substantial processing challenges.

Purpose of the Study:

• To introduce a novel video surveillance management system that optimizes operational efficiency and reduces computational demands.
• To dynamically manage GPU usage and memory allocation for DL-based video surveillance services.
• To improve the balance between GPU memory conservation and DL model reloading latency.

Main Methods:

• Implementation of a two-tiered edge computing architecture (client-server via socket transmission).
• Utilizing federated learning (FL) to train a Long Short-Term Memory (LSTM) network for predicting object appearances.
• Employing a Deep Q-Network (DQN) methodology for dynamically controlling DL model release thresholds, informed by LSTM predictions.

Main Results:

• The proposed system effectively reduces needless GPU usage through a dynamically controlling threshold module.
• Real-time object detection at the primary edge (client side) minimizes data transfer latency.
• The DQN-based dynamic threshold management balances GPU memory conservation with model reloading latency.

Conclusions:

• The novel system significantly improves the efficiency and effective usage of computational resources in DL-based video surveillance.
• This approach enables enhanced security in various domains by optimizing resource allocation and performance.
• Dynamic threshold management using DQN and LSTM predictions offers a superior alternative to static methods.