Centroid based person detection using pixelwise prediction of the position

Implementations of person detection in tracking and counting systems tend towards processing orthogonally captured images on edge computing devices. The ellipse-like shape heads inspired us to predict head centroids determine positions persons images. We the using a fully convolutional network (FCN). combine FCN with simple image operations ensure fast inference detector. experiment size output further decrease time. compare proposed centroid-based detector bounding box-based detectors task terms time performance. propose performance measure which allows quantitative comparison two approaches. For training evaluation detectors, we form original datasets 8000 annotated images, are characterized by high variability lighting conditions, background, quality, elevation profile scenes. an approach simultaneous annotation for both detection. shows best while keeping standards. • A centroid based visual data is proposed. aims at orthogonal scanning scenes variable head–lens distances. Positions sizes coded into gradient ellipses localization maps. Localization maps include that only partially represented scene An efficient transformation introduced. metric box predictions covers false detections as well inaccuracies positioning. New formed detectors.