Cognitive Video Surveillance Management in Hierarchical Edge Computing System with Long Short-Term Memory Model.

Area of Science:

• Computer Science
• Artificial Intelligence
• Edge Computing

Background:

• Deep learning (DL) video surveillance is vital for smart cities, enhancing traffic management and public safety.
• DL models require significant GPU computing and memory resources for object tracking and behavior analysis.
• Current systems face challenges in managing these resource demands efficiently.

Purpose of the Study:

• To develop a novel cognitive video surveillance management (CogVSM) framework.
• To reduce GPU memory consumption in DL-based video surveillance systems.
• To improve the efficiency of object tracking and abnormal behavior detection in edge computing environments.

Main Methods:

• Implementation of a cognitive video surveillance management framework (CogVSM) utilizing a long short-term memory (LSTM) model.
• Forecasting object appearance patterns using time-series data with the LSTM model.
• Adaptive model release control based on LSTM predictions and exponential weighted moving average (EWMA) for dynamic threshold adjustment.

Main Results:

• The LSTM-based CogVSM framework achieved high predictive accuracy with a root-mean-square error of 0.795.
• The proposed framework reduced GPU memory usage by up to 32.1% compared to baseline methods.
• A reduction of 8.9% in GPU memory usage was observed compared to previous works.

Conclusions:

• The CogVSM framework effectively reduces GPU memory requirements for DL-based video surveillance.
• LSTM-based prediction enables adaptive model management, minimizing resource waste.
• The system offers a more efficient and resource-conscious solution for smart city surveillance.