Comparative analysis of multicamera, multilens, fisheye and conventional webcams to address blind spots

Abstract Contemporary security systems integrate object detection algorithms with surveillance cameras to identify and localize objects. While object detection systems are widely used in surveillance applications, their application to screen capture prevention faces practical limitations due to restricted camera field of view. This paper presents a geometric modeling and experimental evaluation framework to mitigate these coverage limitations using cost-effective hardware configurations for preventing smartphone-based data leakage. We derive a mathematical model to determine the optimal number and placement of conventional webcams required to minimize blind regions. In addition, we conduct a systematic comparative analysis of four camera architectures: traditional webcams, multi-camera networks, multilens modules, and fisheye systems. While fisheye lenses provide broader angular coverage, their radial distortion degrades detection performance if not properly addressed. To improve robustness, we introduce a heterogeneous training strategy based on mixed conventional and fisheye image datasets. Experimental results using a YOLOv4-based detector show that the model trained on the mixed dataset achieves 94.6% mAP@0.5, significantly outperforming models trained on single-source datasets in wide-angle deployment scenarios. The findings provide practical guidelines for jointly optimizing camera configuration and model training, demonstrating that combining mixed-dataset learning with wide-angle optics offers a cost-effective and technically robust solution for large-scale corporate display protection.