Multiscale Temporal Feature Fusion for Anomaly Detection in Video Snippets

To enhance the detection of anomalous events that occur across multiple video snippets, this study developed a multiscale temporal feature fusion module. This module leverages multiple 1D convolution kernels of varying sizes to identify local temporal dependencies between instances, leading to better anomaly detection performance. By applying this module to the temporal dimension of video features, it becomes possible to capture local temporal dependencies at different scales.