Efficient School Bullying Detection Based on HPO-TDN

School bullying has always been a source of concern for both schools and parents. With the development of smart cities and the widespread use of cameras inside and outside of schools, it is now possible to identify and stop bullying in real-time via video actions. The traditional approach is to model graphics and timing using sensors, dual-stream networks, or 3D convolutions, and then differentiate between bullying and non-bullying. Problems with these methods include slow inference speed, difficult deployment, and poor practicability. We propose a multi-classification video bully detection method based on Hyper-Parameter Optimization Temporary Difference Networks (HPO-TDN), which combines local and global modeling to fully capture the information in the video. RGB difference is used in local modeling to give more fine-grained feature conditions for 2D convolution. To capture long-term time series structure, cross-segment local modeling is integrated with global modeling. At the same time, the hyper-parameter optimization approach is employed to optimize the parameters. The Tree-structured Parzen Estimator (TPE) algorithm is used to automatically select the momentum parameters in Stochastic Gradient Descent, the learning rate, the dropout, and the number of network layers in the conv3_x and conv4_x layer in the ResNet50 backbone. Finally, a school bully detection model with adaptive network layers is created. The HPO-TDN model improved in accuracy by 1.57% compared to the original TDN model on the five types of school bullying datasets collected.