Fire swin: Video anomaly detection using a hybrid model with convolutional layers, fire module, and swin transformer

There are more than one billion surveillance cameras installed throughout the world that produce enormous volumes of video data on daily basis. Manual search of anomalous events in such huge amounts of data is laborious and extensively time consuming. In contrast, various computer vision methods have been used, and recently deep learning methods have been extensively investigated. However, video anomaly detection in surveillance videos is a complex and challenging task due to varied backgrounds, capturing positions, context, and unexpected events that require high-level semantic understanding and spatiotemporal reasoning. In this paper, we propose an end-to-end novel hybrid model that contains time distributed convolutional layers with a fire module for spatial feature extraction, compression for efficient processing, and a Swin transformer for self-attention mechanism that can capture long-range dependencies and local details. We applied our method to the UCF crime dataset and CUHK Avenue dataset which contains realistic and varied anomalous events with the standard split ratio of used in both UCF crime and CUHK Avenue datasets for training and testing respectively. We demonstrated with extensive experiments that the proposed model achieved 92.20 % accuracy in detecting various anomalies in UCF Crime, and 94.40 % using CUHK dataset, an increase in performance as compared to existing state-of-the-art methods.