A Framework for Integration of Machine Vision with IoT Sensing.

Automated monitoring systems increasingly leverage diverse sensing sources, yet a disconnect often persists between machine vision and IoT sensor pipelines. While IoT sensors provide reliable point measurements and cameras offer rich spatial context, their independent operation limits coherent environmental interpretation. Existing multimodal fusion frameworks frequently lack tight synchronization and efficient cross-modal learning. This paper introduces a unified edge-cloud framework that deeply integrates cameras as active sensing nodes within an IoT network. Our approach features tight time synchronization between visual and IoT data streams and employs cross-modal knowledge distillation to enable efficient model training on resource-constrained edge devices. The system leverages a multi-task learning setup with dynamically adjusted loss weighting, combining architectures like EfficientNet, Vision Transformers, and U-Net derivatives. Validation on environmental monitoring tasks, including classification, segmentation, and anomaly detection, demonstrates the framework's robustness. Experiments deployed on compact edge hardware (Jetson Nano, Coral TPU) achieved 94.8% classification accuracy and 87.6% segmentation quality (mIoU), and they also sustained sub-second inference latency. The results confirm that the proposed synchronized, knowledge-driven fusion yields a more adaptive, context-aware, and deployment-ready sensing solution, significantly advancing the practical integration of machine vision within IoT ecosystems.