Transfer learning model for anomalous event recognition in big video data

Video surveillance faces challenges due to the need for improved anomalous event recognition techniques for human activity recognition. Growing security concerns make standard CCTV systems insufficient because of high monitoring costs and operator exhaustion. Therefore, automated security systems with real-time event recognition are essential. This research introduces a semantic key frame extraction algorithm based on action recognition to minimize frame volume big video data. This approach has not been previously applied with ResNet50, VGG19, EfficientNetB7, and ViT_b16 models for recognizing anomalous events in surveillance videos. The findings demonstrate the effectiveness of this method in achieving high accuracy rates. The proposed method addresses the challenges posed by large volumes of frames generated by surveillance videos, requiring effective processing techniques. A large number of videos from the UCF-Crime dataset were used for proposed model evaluation, including both abnormal and normal videos during the training and testing phase. EfficientNetB7 achieved 86.34% accuracy, VGG19 reached 87.90%, ResNet50 attained 90.46%, and ViT_b16 excelled with 95.87% accuracy. Compared to state-of-the-art models from other studies, the transformer model (ViT_b16) outperformed these algorithms, demonstrating significant improvements in recognizing anomalous events.