TrafficNNode: Low Power Vehicle Sensing Platform for Smart Cities

Automating traffic monitoring and management can materially contribute to the realization of Smart Cities, but this demands detailed, accurate and timely characterization of traffic flows. Current methods employ video capture (high installation and operating costs), pneumatic-tube-based counters (limited detail about vehicle types and often only installed temporarily), or manual data capture (high human cost).We have developed an intelligent, inexpensive, pavement-mountable device capable of collecting information about vehicle types and speeds using an embedded, power-optimized neural network. The device is designed to fit within a raised pavement marker (RPM). RPMs are deployed throughout the world as lane markers. Packaging our sensing technology into RPMs offers the potential to significantly improve the spatial and temporal resolution of traffic flow information across a city.We outline our methodology for energy-optimized machine learning on a small, resource-constrained sensor device. We present the results of our work in terms of accuracy (96% classifying vehicle type and 89% classifying vehicle speed) and battery life (three years).