Recognition of Campus Violence Based on Transformer Model Using Thermal Infrared Videos

To address the issue of campus violence, an algorithm for identifying campus violence based on thermal infrared video data and the Transformer model was proposed in this study. U-Net convolutional network was first used to convert the extracted thermal infrared images from videos into pseudo visible images, enhancing the quality of original images, then the Swin Transformer and the classic Transformer model were sequentially introduced to encode the spatio-temporal features of the sequential images, ultimately achieving the recognition of campus violent behavior. The proposed method exhibits superior performance compared to other commonly used methods, providing strong support for enhancing the intelligence of campus safety management and promoting the development of smart education.