People detection using time-of-flight camera

We study the problem of detecting people using time-of-flight camera images, i.e. depth images. These detections are used to measure the accurate flow of people through a security gate. This is a major problem in the case of access control, such as in a metro station or an airport. The detection of people from depth images is challenging because of the variety of possible scenarios and implied shapes. The scenarios can involve people in a wheelchair, pushing bikes, or carrying suitcases, bags, etc. We show that the detection can be processed by filtering and classifying regions of interest around the local maxima of the depth images. We then propose to consolidate the detections across a scenario with multi-object tracking. The missing and extra people are measured for the performance analysis. Our algorithm is developed within Matlab, and trained on a dataset of top-view depth images of people crossing a metro access gate. Extensive validation demonstrates its effectiveness. Detection rates on stand-alone images larger than 80% are typically achieved with fewer than 10% false positives. Measurement of the people flow is correct in 90.91% of the scenarios.