Key points are not available for this paper at this time.
Integrating vision-based technologies into distributed sensor domains offers unprecedented potential for data collection. However, it raises privacy concerns over the incredible amount of extra information inadvertently carried by the video stream. On the other hand, the advent of tiny machine learning models running on edge devices with embedded GPUs/TPUs is revolutionizing computer vision and real-time tracking systems, enabling the local execution of computationally demanding tasks traditionally performed in the cloud. This study focuses on developing and characterizing vision-based virtual sensors capable of processing data from a local camera source to provide real-time measures of relevant metrics without storing or transmitting any video stream. The main advantages of vision-based virtual sensors running on the edge are data protection, reduced communication cost, and reduced detection latency. In addition, we propose a dynamic inference power manager (DIPM), based on adaptive frame rate, that allows us to explore the trade-off between power consumption and accuracy. Experimental results conducted on a real hardware platform show that the proposed virtual sensor, equipped with DIPM, can save up to 40% of the processing energy with a reduction of tracking accuracy lower than 10%, while retaining the privacy preservation benefits of virtual sensors.
Kania et al. (Mon,) studied this question.