Real-world Anomaly Detection in Surveillance Videos using YOLO-World and BiLSTM Framework

Abstract

Detecting suspicious activities using Closed-Circuit Television (CCTV) is essential for improving security in public and semi-public spaces like schools, colleges, residential areas, parks, and workplaces. Conventional surveillance systems depend largely on constant human oversight from security staff, which can become ineffective and prone to mistakes when trying to monitor several camera feeds at once. This frequently leads to slow reactions or overlooking suspicious or unusual behaviors. This project suggests an automated suspicious activity detection system that uses deep learning techniques implemented using TensorFlow and Python in order to overcome these restrictions. A human-focused open-vocabulary detector called YOLO-World and BiLSTM Framework is employed. After suppressing the background noise, BiLSTM is employed to comprehend temporal behavior. In order to spot unusual or suspect human activity in real time, the system continuously examines live CCTV video streams. Instant notifications are created and transmitted to authorized personnel upon detection of such events, facilitating quicker and more informed decision-making. By automating the monitoring process and reducing reliance on humans, the suggested method greatly lessens the cognitive strain on security officers. It guarantees continuous observation without weariness, increases accuracy, and speeds up response times. The suggested method provides a more effective, scalable, and intelligent security mechanism than traditional surveillance systems. The technology helps create safer and more secure environments, which benefits society as a whole by facilitating prompt intervention and proactive threat identification.