Anomaly detection in surveillance videos based on H265 and deep learning

• Published in: International Journal of Advanced Technology and Engineering Exploration Vol. 9; no. 92; p. 910
• Main Authors: Abbas, Zainab K, Al-Ani, Ayad A
• Format: Journal Article
• Language: English
• Published: Bhopal Accent Social and Welfare Society 31.07.2022
• Subjects: Anomalies; Artificial neural networks; Closed circuit television; Crime; Datasets; Deep learning; Feature extraction; Frames (data processing); Human activity recognition; Machine learning; Sampling; Security systems; Surveillance; Video compression
• ISSN: 2394-5443; 2394-7454
• DOI: 10.19101/IJATEE.2021.875907

This paper discusses anomaly detection, which is one of the most well-known applications of human activity recognition. Due to the ever-increasing activities posing risks ranging from planned aggression to harm caused by an accident, providing security to an individual is a major issue in any community today. Traditional closed-circuit television does not suffice since it necessitates a human being to be awake and always watch the cameras, which is costly. This necessitates the creation of an automated security system that detects anomalous activity in real-time and provides rapid assistance to victims. However, identifying activity from long surveillance footage takes time. Hence, in this research, we study the effect of the down-sampling concept of the challenging database, namely the university of central Florida (UCF Crime) using high efficiency video coding (HEVC)-H265 before feeding them into the anomaly detection system. This step reduced the size of the data, making it easier to store and transfer, and highlights the unique properties of each video clip. In the proposed work, first, we are down-sampling each video’s frame into half by using H265 on the fast forward moving picture experts group (FFMPEG) platform, and then spatiotemporal features are extracted from a series of frames (frame level) using a pre-trained convolutional neural network (CNN) called Resnet50, then to boost the feature we are combining the features of every 15 video frames to generate a new feature vector that will be fed into the classifier model. The values of the new feature vectors represent the summation of the values of the original feature vectors obtained from Resnet50. Finally, the features obtained from a series of frames are fed to the bidirectional long short-term memory (BiLSTM) model, to classify the video as normal or abnormal. We conducted comprehensive tests on a different benchmark dataset for anomaly detection to verify the proposed framework's functionality in complex surveillance scenarios. The numerical results were carried out on the UCF crime dataset, with the proposed approach achieving an area under curve (AUC) score of 90.16% on the database's test set.