Low-Resolution Overhead Thermal Tripwire for Occupancy Estimation

How Low Can You Go?

“…We investigate the effect of different camera configurations by systematically manipulating the camera resolution and frame rate and assessing their impact on occupancy detection. We use a generic processing pipeline that has been inspired by state-of-the-art systems for occupancy counting using thermal array sensor data [12,13]. Indeed, the main contribution of our work is the systematic evaluation of camera configurations and the resulting insights about trade-offs in the use of thermal array sensors with differing resolutions, not the processing of thermal images itself.…”

“…As our primary dataset we use the Thermal Images for Door-based Occupancy Sensing (TIDOS) dataset [13], which was collected to provide a privacy-preserving method to estimate the amount of people indoors through thermal images. TIDOS contains 118 208 frames (123 minutes and 8 seconds of video).…”

“…Low-resolution thermal array sensors are emerging as a promising solution for occupancy detection and counting in residential homes as the sensors are inexpensive, have low deployment cost, and are both energyefficient and non-intrusive to occupants [6,7,13,32,46]. A significant challenge with low-resolution sensors, however, is that their performance is highly sensitive to the overall resolution, frame rate, and the field-of-view of the camera [46].…”

“…Tyndall et al [46] consider 4 × 16 sensor arrays whereas Cokbas at al. [13] rely on 32 × 16 sensor arrays. Understanding the different performance trade-offs of such systems is essential for generalizing these systems and offering solutions that can be adopted to different building environments.…”

“…We conduct our analysis through systematic benchmarks using image manipulations. We consider two datasets, an open-source dataset with tripwire-triggered thermal images (TIDOS [13]) and a proprietary dataset that is used to validate the generality of the findings and to study the effect of differences in the camera's fields-of-view. The results of our evaluation show that even cameras with a 4 × 2 resolution -significantly lower than what has been used in previous research -can support reliable detection, as long as the frame rate is at least 4 frames per second.…”