A fast object detector based on high-order gradients and Gaussian process regression for UAV images

Unmanned aerial vehicles (UAVs) acquire images characterized by an exceptional level of detail, which calls for processing and analysis methods capable of efficiently exploiting their rich information content. In particular, the detection of specific classes of objects (e.g. cars, roofs) represents an important but challenging task for these images. Most of the related literature has aimed at proposing methods capable of providing satisfactory detection accuracies. However, they typically refer to a specific class of objects and give little attention to the processing time. In this work, we present a novel and fast methodological alternative. In addition to being particularly fast, the proposed method is a general detection approach that can be customized to any class of objects after an opportune training phase. It consists of the design of a non-linear filter that combines image gradient features at different orders and Gaussian process (GP) modelling. High-order image gradients permit one to capture detailed information regarding the structure of the investigated class of objects. The GP model fed with high-order gradients yields an estimate of the presence of the object of interest for a given position of the sliding window within the image. Two separate sets of experiments were conducted, each aiming at assessing the proposed method to detect a given class of objects common in urban scenarios, namely vehicles and solar panels, respectively. Results on real UAV georeferenced images characterized by 2 cm resolution and by three channels (red, green, and blue (RGB)) are provided and discussed, showing particularly interesting properties of the detector.