An Integrated CNN-BiLSTM-Transformer Framework for Improved Anomaly Detection Using Surveillance Videos

Surveillance cameras have been widely used for monitoring in both private and public sectors as a security measure. Close Circuits Television (CCTV) Cameras are widely used to surveillance normal and anomalous incidents. Real-world anomaly detection in complex environment situations is a significant challenge due to its diverse nature manually reviewing this vast data is difficult, time-consuming, and prone to errors. Human operators may overlook crucial information due to failure, which reduces the accuracy of manual monitoring. This highlights the need for automated solutions to improve accuracy and efficiency We proposed a Bidirectional Motion Temporal (BiMT) model to address this issue by coupling a pre-trained model of deep neural networks. This system utilizes a Convolutional Neural Network (CNN) to extract the best spatial key features from the video stream, which are converted into time series data and coupled with Recurrent Neural Network (RNN) deep learning model Bidirectional Long Short-Term Memory (BiLSTM) extract temporal key features, and Multi-Head Self-Attention (MHSA) allows for the detection of short-term frame correlations. In addition, created a custom transformer encoder with relative positional embedding that improves the detection and recognizes long-range frame dependencies. Furthermore, categorical focal loss (CFL) is used during training to prioritize essential features and further enhance the model's precision. The more sophisticated ConvNeXtv2 model produced average AUC scores of 98.6%, 86.5%, and 83.5% for the UCF-Crime, UBI-Fight, and RAD datasets respectively.