Detection of Abnormal Activities in a Crowd Video Surveillance using Contextual Information

This paper introduces an enhanced method for detecting abnormal behavior in surveillance videos using contextual information. It builds upon previous research in the field, incorporating Convolution Neural Networks (CNNs), Recurrent Neural Networks (RNN), Auto-encoder architectures, and Vision Transformer (ViT) models with self-attention mechanisms. This self-attention feature effectively captures contextual information from video frames, surpassing existing state-of-the-art models in both accuracy and robustness. Furthermore, a comparative analysis will be presented, comparing the results obtained by our model with those of other cutting-edge approaches. In our study, we enhance the generalizability of our model by combining two datasets encompassing various violent scenarios. These datasets include the "fight-detection-surv-dataset". and the "Video Fight Detection Dataset" which consists of regular surveillance camera videos and samples from YouTube videos. By leveraging these diverse sources, we aim to train our model effectively on multiple instances of violent behavior.