Automatic Computer Detection and Power Estimation in Indoor Environments from Imagery

In this paper, we describe an automated technique for estimating power consumption by computers in buildings by processing visual imagery collected during a walkthrough of the building. This is an important problem since desktop and laptops are the largest contributors to electricity consumption in most large commercial buildings. The images used for estimation are obtained by a backpack equipped with a suite of sensors such as laser scanners, cameras, and an IMU, carried by a human operator walking at normal speed inside buildings. In addition, the operator carries a handheld infrared (IR) camera to take pictures of the CPU box of the desktops. We take a two step approach to this problem. First, we develop a technique based on convolutional neural networks to detect and count computers, which results in 90% accuracy as tested on a three story building with over 100 machines. Second, we develop an SVM based power estimation algorithm for computers using a handheld IR camera which captures both IR and visible light imagery simultaneously. The average power estimation error over 101 computers is around 8%. Combining these two algorithms, it is possible to accurately power consumption due to computers in commercial buildings.