Detecting African hoofed animals in aerial imagery using convolutional neural network

Small unmanned aerial vehicles (UAVs) applications had erupted in many fields including conservation management. Automatic object detection methods for such aerial imagery were in high demand to facilitate more efficient and economical wildlife management and research. This paper aimed to detect hoofed animals in aerial images taken from a quadrotor in Southern Africa. Objects captured in this way were small both in absolute pixels and from an object-to-image ratio point of view, which were not perfectly suit for general purposed object detectors. We proposed a method based on the iconic Faster R-CNN framework with atrous convolution layers in order to keep the spatial resolution to detect small objects. A good choice of anchors was of prime importance in detecting small objects. The proposed method was compared to RetinaNet as a baseline, the performance of the atrous filters for feature extraction was proven to be outstanding in this case by comparing to Feature Pyramid Network, a literal connection approach. The results outperformed both frameworks by a great margin. Detecting African hoofed animals in aerial imagery using convolutional neural network