Real-Time Object Detection with Automatic Switching between Single-Board Computers and the Cloud

We present a wireless real-time object detection system utilizing single-board devices, cloud computing platforms and web-streaming. Currently, most inference applications stat- ically perform tasks either on local machines or remote cloud servers. However, devices connected through cellular technolo- gies face volatile network conditions, compromising detection performance. Furthermore, while the limited computing power of single-board computers degrade detection correctness, exces- sive power consumption of machine learning models used for inference reduces operation time. In this paper, we propose a dynamic system that monitors embedded device’s wireless link quality and battery level to decide on detecting objects locally or remotely. The experimental results show that our dynamic offloading approach could reduce devices’ energy usage while achieving high accuracy, real-time object detection. Index Terms—Machine learning, WebRTC, object detection