Video-based abnormal human behaviour detection for video forensics

Detection of abnormal human behaviour in video footage is vital for video forensics, aiding in identifying suspicious activities. Deep learning techniques, particularly convolutional neural networks (CNNs), have demonstrated promising results in this area, by using CNNs for feature extraction. However, existing methods, especially 3D CNN models employed for detecting abnormal actions, focus solely on local information, limiting their capacity to capture broader patterns or dependencies extending across larger spatial and temporal scales. This paper proposes a new method that integrates a 3D convolution attention block to capture pertinent features in abnormal actions, enhancing digital forensics for crime investigation. By prioritizing the extraction of dependencies between dimensions, our approach captures intra-spatial details within frames and inter-temporal connections across sequences, providing a comprehensive global view. Evaluation on RLVS and UCF-crime anomaly detection datasets demonstrates consistent improvements in AUC performance by 2% to 3% compared to conventional methods, showcasing the efficacy of integrating 3D-SCT.