Exploiting On-Device Image Classification for Energy Efficiency in Ambient-Aware Systems

Ambient-aware applications need to know what objects are in the environment. Although video data contains this information, analyzing it is a challenge esp. on portable devices that are constrained in energy and storage. A naive solution is to sample and stream video to the cloud, where advanced algorithms can be used for analysis. However, this increases communication energy costs, making this approach impractical. In this article, we show how to reduce energy in such systems by employing simple on-device computations. In particular, we use a low-complexity feature-based image classifier to filter out unnecessary frames from video. To lower the processing energy and sustain a high throughput, we propose a hierarchically pipelined hardware architecture for the image classifier. Based on synthesis results from an ASIC in a 45 nm SOI process, we demonstrate that the classifier can achieve minimum-energy operation at a frame rate of 12 fps, while consuming only 3 mJ of energy per frame. Using a prototype system, we estimate about 70 % reduction in communication energy when 5 % of frames are interesting in a video stream.