ADAPTIVE DEEP LEARNING FRAMEWORK FOR REALTIME SUSPICIOUS ACTIVITY RECOGNITION IN SMART SURVEILLANCE SYSTEMS

Abstract

With the growing adoption of smart surveillance systems in public and private environments, the need for intelligent automated suspicious activity detection has become increasingly critical. Traditional video monitoring methods depend heavily on human operators and rule-based algorithms, leading to high false alarm rates and limited scalability. This research proposes an adaptive deep learning framework capable of recognizing suspicious human activities in real time. The system integrates Convolutional Neural Networks (CNNs) for spatial feature extraction and Long Short-Term Memory (LSTM) networks for temporal sequence modeling, allowing the detection of complex behavioral patterns over time. To ensure adaptability, the model dynamically adjusts to changing lighting and environmental conditions using realtime feedback and transfer learning. Experiments conducted on benchmark surveillance datasets demonstrate that the proposed framework achieves superior accuracy, robustness, and responsiveness compared to conventional methods. The approach can significantly enhance situational awareness and strengthen security operations in smart cities and critical infrastructures.