Video anomaly detection using transformer
Ranjeet Madhav, Siddharth Bothra, Sanjay Kumar · 2025
This research paper delves into the utilization of TimeSformer, a transformer-based architecture tailored for processing temporal data, in the realms of video classification and anomaly detection. Building upon TimeSformer&s;s prowess in recording long-range temporal dependencies efficiently, we suggest a novel framework that integrates TimeSformer for feature extraction alongside an attention mechanism comprising dilated convolution and self-attention layers, enabling the model to record both short and long-range temporal dependencies effectively and anomaly detection using Robust Temporal Feature Magnitude Learning (RTFM). Through validation on the open source ShanghaiTech Campus dataset, our offered framework demonstrates competitive performance compared to existing methods, underscoring its efficacy in surveillance video analysis.