DEVTrV2: Enhanced Data-Efficient Video Transformer For Violence Detection
Almamon Rasool Abdali, Ammar Abdullah Aggar · 2022 7th International Conference on Image, Vision and Computing (ICIVC) · 2022
In smart surveillance systems, event classification and detection are essential parts, on the other hand, violent event recognition is one of the most important key elements in that systems. The use of the transformer network for video action recognition has achieved very high results, On the other hand, the transformer network needs a large amount of data to gain good accuracy therefore, a new family of video transformer networks called Data-efficient video transformer (DEVTr) came to address these problems and shows that the use of the pre-trained convolution neural network in the embedding stage can improve accuracy with the use of a small dataset for the given task. In this study, we extend the data-efficient video transformer (DEVTr) for better event detection within a small dataset and with low hardware resources. We also introduced two new video data augmentations methods (random erase from frames and frame position shifting with blurring). The model achieved 98.25% accuracy on the Real-life violence dataset (RLVS), an accuracy of 96% on the NTU CCTV-fight dataset, and accuracy of 91.803% on the UBI-Fight dataset. A comparison with previous techniques illustrated that the proposed methods provide the best result among the other research for violence event detection within a small dataset and low hardware resources.