Unobtrusive People Identification and Tracking Using Radar Point Clouds

• Published in: IEEE sensors letters Vol. 7; no. 12; pp. 1 - 4
• Main Authors: Chowdhury, Arijit, Pattnaik, Naibedya, Ray, Arindam, Chakravarty, Soumya, Chakravarty, Tapas, Pal, Arpan
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.12.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Continuous radiation; Feature extraction; frequency modulated continuous wave (FMCW) radar; gait; person identification; Point cloud compression; point clouds; PointNet; Radar; Radar tracking; Sensor applications; Sensors; Smart buildings; Surface treatment; Target tracking
• ISSN: 2475-1472; 2475-1472
• DOI: 10.1109/LSENS.2023.3328794

Identification of people in closed spaces is an indispensable requirement in modern smart home spaces. Existing recognition methods that utilize vision sensors, such as cameras, cannot be used for this purpose because of their privacy-invasive sensing characteristics. In this letter, we propose a method for unobtrusive identification and tracking of people by capturing their unique gait pattern in a closed space using point clouds generated from commercially available frequency modulated continuous wave radars. We primarily focus on handling the nonlinearity due to the variation of the subject's distance from the radar by augmenting the point clouds with novel height surface maps that are generated individually for every person. We build a two-level feature generation system on top of these point clouds to uniquely identify them. We also attempt identification using a blend of these height surface maps and existing point cloud processing architectures, such as PointNet and PointNet++. The average precision and recall for all seven subjects tested were <inline-formula><tex-math notation="LaTeX">79.28\%(\pm 4.9)</tex-math></inline-formula> and <inline-formula><tex-math notation="LaTeX">80.23\%(\pm 9.8)</tex-math></inline-formula>. Finally, the proposed method augments the height surface maps with the PointNet architecture and utilizes majority voting scheme for people identification. It provides an accuracy above 90%, which indicates the efficiency of our implemented solution.