Presentation title

Low noise and low power techniques for tagless indoor human localization and identification using capacitive sensors

Authors

Mihai Lazarescu and Luciano Lavagno

Institution(s)

Politecnico di Torino, Italy

Presentation type

Technical presentation

Abstract

Indoor human detection, localization and identification is at the base of many automation and monitoring systems. Capacitive sensors do not require the persons to wear tags or to otherwise interact with the system. Moreover, sensors working in load-mode can be small, inexpensive, low power, and easy to install and to operate in new building or to retrofit existing ones. However, especially when sensing at ranges much longer than their size (the square of the plate area), the measurement accuracy of capacitive sensors can be significantly influenced by several environmental factors and types of noise. On the other hand, their sensitivity to noise is often complementary to other types of sensors (such as infrared and ultrasound), which makes capacitive sensors good candidates for measurement systems based on sensor fusion. Indoor activity of persons (e.g., for assisted living) often includes long periods of low or absence of movement (e.g., when watching TV or sleeping). Such periods can extend for several hours during which the capacitance measurements can be heavily influenced by very low frequency noise, such as 1/f. Its effect often appears as a drift of the measurements, which can exceed several times the amplitude of the signal. Signal filtering has limited effect, since human activities can have a large bandwidth, from a few Hz (associated with walking) down to near DC (for long periods of physical inactivity). Besides, digital filters for very low frequencies can be very resource-demanding for small embedded devices. To address such problems, we present a capacitance measurement technique that shows a very good rejection of low frequency noise, including drifts. The technique can be used for simple, single-plate transducer layouts. It is based on differential measurements combining analog and digital signal processing. To assess its effectiveness, we present simulation results that show the noise rejection ratio for various noise frequencies, compared to single-ended measurements. We also show results of indoor person identification using neural networks, based on capacitive response of the human body at different measurement frequencies. We show how training and inferring are influenced by the amount of noise and the structure of the neural network. We also measure the resource and energy consumption of neural network implementations on low power FPGAs.

---

Additional material

• Presentation slides: [pdf]

For more details on this presentation please click the button below: