Comparative Analysis of YOLOv3, MobileNet-SSD, and EfficientDet for Real-Time Person Detection in Low-Resolution Images

Object detection using low-resolution cameras is increasingly important for resource constrained and low-cost systems. This study evaluates the performance of three pre-trained machine learning models from the COCO dataset YOLOv3, MobileNet-SSD, and EfficientDet for human detection at three image resolutions: 160×120 (QQVGA), 320×240 (QVGA), and 640×480 (VGA). Low resolutions pose challenges due to limited visual and spatial details, which can reduce detection accuracy. Results show that YOLOv3 achieves the highest accuracy and consistency across all resolutions but demands higher computational resources, reflected in increased CPU usage, higher latency, and lower frame rates (FPS). MobileNet-SSD provides a balanced trade-off between accuracy and efficiency, making it suitable for real-time applications in power-constrained environments. EfficientDet, although the most lightweight in computation, shows lower detection accuracy. This study emphasizes empirical evaluation under real-world deployment constraints, following the zero-shot performance approach, where pre-trained models are applied directly to low-resolution edge data without retraining. The findings offer practical insights into the trade-offs among accuracy, efficiency, and latency, guiding model selection for IoT and edge computing applications in resource-limited settings. This work contributes to understanding suitable model choices for low-resolution human detection in real-world scenarios.