A Cloud-Based Deep Learning Framework for Real-Time Anomaly Detection in Public Surveillance Videos Using VGG16-LSTM

The purpose of this research is to design and train a deep learning-based surveillance network that can detect anomalous and suspicious behavior in real time using the Daily Crime Surveillance and Safety System (DCSASS). The system will enhance safety on the streets and enable detection of threats ahead through automation of anomaly detection of the video feeds. The methodology is systematic and organized and starts by gathering and authenticating video data of a high risk activity of 13 high risk classes of activities such as assault, robbery, arson and vandalism. An effective preprocessing pipeline was developed, including a Gaussian blur filter and brightness-contrast normalization to improve clarity of the frame and reduce noise. The videos were then split into frames and annotated by using official labels and separated into training and validation sets using stratified sampling to ensure class balance. A hybrid architecture was followed whereby the VGG16 convolutional neural network that was used as the frozen spatial feature extractor was integrated with the Long Short-Term Memory (LSTM) layer to learn temporal interrelations. The time intervals of the input feature vectors were converted to a repetition, simulating the motion context. The performance, such as the accuracy, loss, precision and recall, were modelled using the standard metrics. The proposed VGG16-LSTM model presented an accuracy of 95.05, low loss of 0.126, precision of 95.46 and recall of 93.11 indicating that this model is very reliable when it comes to detecting anomalies. In addition, object detection analysis yielded a precision of 85.7% and mAP 0.5 of 71.7%. The validity and scalability of the suggested framework were validated in comparison to the other existing models, including ResNet-50 and an earlier hybrid model.