An innovative deep learning based approach for anomaly detection in intelligent video surveillance

Abstract

Nowadays, anomaly detection has gained vital importance as security is a major concern everywhere. This work focuses on developing an intelligent video surveillance system capable of detecting anomalous activities in videos, utilizing the UCF Crime dataset as the primary source. The proposed model employed a multistage method uniting the convolutional neural networks (CNN) and long short-term memory (LSTM) networks. In the proposed approach, video frames serve as input to the CNN, which processes them to extract key features. These features are then passed to an LSTM network to capture temporal dependencies and identify anomalous events over time. This CNN-LSTM architecture successfully detects twelve distinct types of anomalous activities: abuse, arrest, arson, assault, burglary, explosion, fight, road accident, robbery, stealing, shoplifting, and vandalism. The dataset is divided into portions for training, testing, and validation, along with cross-validation to ensure model generalization. The system achieves an accuracy of 98.6%, reflecting a significant improvement of 4-5% over existing systems. This demonstrates the robustness of the proposed method in detecting anomalous behavior in video data.