A Multimodal Synthetic Dataset for Multi-Camera Human Detection and Occlusion Analysis in Indoor Environments

Abstract

Synthetic data has become an essential component in modern computer vision and robotics research, particularly in applications where collecting large, diverse, and fully annotated real world datasets are impractical or impossible. This study presents a high resolution, multimodal, and multi camera synthetic dataset specifically developed for human detection and partial visibility analysis in indoor environments. Two distinct scenarios were designed within NVIDIA Isaac Sim, involving a total of eleven cameras positioned across separate residential spaces. The use of 120 FPS animations and synchronized multimodal annotation generation enabled detailed capture of human motion, occlusion, and scene variability. Diverse human models, including variations in appearance, clothing, accessories, and behavior, were incorporated to replicate real world heterogeneity and challenge vision algorithms under complex conditions. Despite the benefits of precise annotation and full environmental control, the study also revealed clear constraints related to computational load and real time simulation performance, particularly when generating dense annotation sets. The resulting dataset nonetheless provides a rich and comprehensive foundation for research in human tracking, multi camera fusion, behavior understanding, and security oriented computer vision systems. The expanded analysis concludes that synthetic data, when produced through high fidelity simulation workflows, offers a practical, ethical, and scalable alternative that can significantly advance both methodological and applied research. © 2026 IEEE.