The increasing demand for intelligent surveillance due to rising crime rates has driven the development of automated video monitoring in Vietnam. This paper presents an enhanced long-term recurrent convolutional network (LRCN) model that integrates convolutional neural networks (CNNs) for spatial feature extraction and long short-term memory (LSTM) for temporal analysis, enabling real-time detection of suspicious activities like theft and intrusion. The system achieves over 85% accuracy on KTH Action (Laptev and Caputo in Recognition of human actions—Action Database) and UCF-crime (Sultani et al. in UCF-crime dataset) datasets, demonstrating its effectiveness in real-world applications. Designed to be scalable and cost-efficient, it addresses infrastructure limitations and is adaptable for residential, commercial, and large-scale security while reducing dependence on human monitoring. This research enhances security by providing a flexible, AI-driven solution for modern surveillance challenges.