Photorealistic Synthetic Crowds Simulation in Indoor environments (PSCS-I): a novel synthetic dataset for realistic simulation of crowd panic and violence behaviors

Stefanos Pasios, Konstantinos Gkountakos, Konstantinos Ioannidis, Theodora Tsikrika, Stefanos Vrochidis, Ioannis Kompatsiaris

Abstract: Crowd analysis is a critical research field in computer vision that focuses on understanding and predicting crowd behaviors with applications extending to public safety, urban planning, and event management. However, the scarcity of large-scale, diverse, annotated datasets that can be employed for benchmarking crowd analysis-related methods has led to a stagnation in research, with many approaches already achieving near-perfect accuracy on the available datasets. In this work, we propose a novel dataset for Photorealistic Synthetic Crowds Simulation in Indoor environments (PSCS-I), generated with the state-of-the-art rendering capabilities of Unreal Engine 5, that addresses these challenges. Our dataset contains 647 photorealistic videos of long duration with high-quality annotations that support surveillance monitoring systems for crowd panic and violence behaviors detection tasks. Based on this dataset, we performed a number of experiments with the current state-of-the-art crowd behavior analysis-related method, with the results demonstrating that the dataset can pose a significant challenge to multiclass behavior detection and that it outperforms the current crowd behavior analysis benchmarks on cross-dataset evaluation for abnormal crowd behavior detection. The dataset will be made publicly available upon acceptance at https://m4d.iti.gr/results/.

External IDs:doi:10.1109/access.2025.3636727