Data generation via diffusion models for crowd anomaly detection

Crowd analysis is a critical aspect of public security and video surveillance. One of the primary challenges in developing effective crowd anomaly detectors is the lack of comprehensive training data. To address this issue, we investigate using synthetic data to enhance training for anomaly detection in crowded environments by generating a dataset of synthetic videos using two open-source diffusion models. Each synthetic video depicts typical crowded scenes that may be either normal or anomalous. To assess the effectiveness of our approach, we compare the model’s performance across three training scenarios: using only real videos, only synthetic videos, and a combination of both. This preliminary analysis highlights the potential of data generated via diffusion models to improve crowd anomaly detectors’ stability and classification capabilities.