Spatial Resolution-Independent CNN-Based Person Detection in Agricultural Image Data

Advanced object detectors based on Convolutional Neural Networks (CNNs) offer high detection rates for many application scenarios but only within their respective training, validation and test data. Recent studies show that such methods provide a limited generalization ability for unknown data, even for small image modifications including a limited scale invariance. Reliable person detection with aerial robots (Unmanned Aerial Vehicles, UAVs) is an essential task to fulfill high security requirements or to support robot control, communication, and human-robot interaction. Particularly in an agricultural context, persons need to be detected from a long distance and a high altitude to allow the UAV an adequate and timely response. While UAVs are able to produce high resolution images that enable the detection of persons from a longer distance, typical CNN input layer sizes are comparably small. The inevitable scaling of images to match the input-layer size can lead to a further reduction in person sizes. We investigate the reliability of different YOLOv3 architectures for person detection in regard to those input-scaling effects. The popular VisDrone data set with its varying image resolutions and relatively small depiction of humans is used as well as high resolution UAV images from an agricultural data set. To overcome the scaling problem, an algorithm is presented for segmenting high resolution images in overlapping tiles that match the input-layer size. The number and overlap of the tiles are dynamically determined based on the image resolution. It is shown that the detection rate of very small persons in high resolution images can be improved using this tiling approach.