Detection of people from time-of-flight depth images using a cell-tracking methodology

Abstract

This paper describes a method of detection and tracking of people using depth images captured by a Time-of-Flight (ToF) camera, such as those obtained with a Microsoft Kinect V2. Key advantages of this approach are that the identity of people is not disclosed, and the system can operate in low-light conditions. The automated approach developed here is inspired by cell-tracking methods as implemented in the well-known biomedical imaging software called CellProfiler. Our approach involves significant preprocessing of the depth images by a combination of adaptive center weighted median filtering and iterative inpainting. The next step is detection of each person’s head using depth local minima information. The classification of each person is typically possible using evidence of shoulder depth information assisted by Laplacian of Gaussian (LoG) based matched filtering. After some additional processing and blob analysis, further quantification and monitoring of people is done using a multi-object tracking system based on Kalman filtering. The main applications are in the general area of collection of statistical information in smart building entrances/exits for security and commercial use.