Recognizing the Age, Gender, and Mobility of Pedestrians in Smart Cities using a CNN-BGRU on Thermal Images

Abstract

In order to extend ambient assisted living technologies for semi-autonomous people from smart homes to smart cities, it is necessary to recognize vulnerable people in the city. Gait-based approaches have been used to perform soft biometrics recognition, such as age or gender. However, most of these approaches rely on the use of RGB cameras, which can lead to serious privacy issues in the city. In this paper, we present an approach to perform age, gender, and mobility recognition on pedestrians in the city using thermal cameras, thus ensuring a high degree of privacy for the citizens. This work is a follow-up to an initial shallow CNN-based approach. A deeper CNN is optimized and used to perform age, gender, and mobility recognition on a single image before extending this approach to a sequence of images using a CNN-BGRU for age and gender recognition. The results obtained using these approaches are presented before discussing the social implications of a real-time gait-based soft biometrics recognition system in the city using thermal cameras. The system has achieved an accuracy of 77.00%, 78.46%, and 94.77% respectively for age, gender and mobility recognition using unprocessed 160x120 thermal images containing a majority of partial pedestrians silhouettes in the city.