A regression-based radar-mote system for people counting

People counting is key to a diverse set of sensing applications. In this paper, we design a mote-scale event-driven solution that uses a low-power pulsed radar to estimate the number of people within the ∼10m radial range of the radar. In contrast to extant solutions, most of which use computer vision, our solution is light-weight and private. It also better tolerates the presence of obstacles that partially or fully impair line of sight; this is achieved by accounting for “small” indirect radio reflections via joint time-frequency domain features. The counter itself is realized using Support Vector Regression; the regression map is learned from a medium sized dataset of 0–∼40 people in various indoor room settings. 10-fold cross validation of our counter yields a mean absolute error of 2.17 between the estimated count and the ground truth and a correlation coefficient of 0.97.We compare the performance of our solution with baseline counters.