Search

Monitoring the number of people in various spaces of a building is important for optimizing space usage, assisting with public safety, and saving energy. Diverse approaches have been developed for different end goals, from ID card readers for space management, to surveillance cameras for security, to CO2 sensing for HVAC control. In the last few years, fisheye cameras mounted overhead have become the sensing modality of choice because they offer large-area coverage and significantly-reduced occlusions but research efforts are still nascent. In this paper, we provide an overview of recent research efforts in this area and propose one new direction. First, we identify benefits and challenges related to inference from top-view fisheye images, and summarize key public datasets. Then, we review efforts in algorithm development for detecting people from a single fisheye frame and from a group of sequential frames. Finally, we focus on counting people indoors. While this is straightforward for a single camera, when multiple cameras are used to monitor a space, person re-identification is needed to avoid overcounting. We describe a framework for people counting using two cameras and demonstrate its effectiveness in a large classroom for location-based person re-identification. To support people counting in even larger spaces, we propose two new person re-identification algorithms using N > 2 overhead fisheye cameras. We provide ample experimental results throughout the paper.

Person re-identification (PRID) has been thoroughly researched in typical surveillance scenarios where various scenes are monitored by side-mounted, rectilinear-lens cameras. To date, few methods have been proposed for fisheye cameras mounted overhead and their performance is lacking. In order to close this performance gap, we propose a multi-feature framework for fisheye PRID where we combine deep-learning, color-based and location-based features by means of novel feature fusion. We evaluate the performance of our framework for various feature combinations on FRIDA, a public fisheye PRID dataset. The results demonstrate that our multi-feature approach outperforms recent appearance-based deep-learning methods by almost 18% points and location-based methods by almost 3% points in matching accuracy. We also demonstrate the potential application of the proposed PRID framework to people counting in large, crowded indoor spaces.