Low-cost face tracking using ONVIF-based surveillance cameras

In this contribution, we investigate consumer-grade surveillance cameras for use as face-tracking cameras in laboratory settings. For the face detection, we compare three algorithms that work out-of-the-box in OpenCV, using pre-trained models, namely YOLOv3, Caffe, and Haar Cascades. For Caffe, this is the first true performance measurement in literature to our knowledge. On a test set of 316 images we created, half without face, Caffe performed best overall, reaching a classification accuracy of 98.1% at a framerate of 30 fps. Using this model, we then constructed a simple two-point controller, to control a low-cost surveillance camera using the open ONVIF standard. Despite some problems due to the simple nature of the controller, the setup worked reasonably well, following a person moving around in a cobot workspace at speeds no faster than 8.5°/s. We conclude that surveillance cameras possessing the open ONVIF standard, paired with a standard face detector implemented out-of-the-box in OpenCV, can be a feasible low-cost alternative to more expensive face tracking cameras.