[Paper] Deep Learning based Hierarchical Object Detection System Adopting a Depth Correction Scheme for High-Resolution Aerial Images Towards Realization of Autonomous UAV Flight

Abstract

We propose deep learning based Hierarchical Object Detection System (HODS) adopting a depth correction scheme designed to improve detection performance in high-resolution aerial images obtained by UAVs. This system consists of a hierarchical two-stage inference unit such as Global Detecting Unit (GDU) and Local Detecting Unit (LDU) for detecting large and small objects respectively, while the Small-object Collecting Unit (SCU) component creates a reconstruction image for local detection. The high-resolution aerial images are first downsampled, and then coarse inference is performed by the GDU to detect large objects and small object candidates. Next, based on the distribution of small object candidates, the SCU collects rectangular areas where small objects are located and creates a reconstruction image with a new normalization approach, the depth correction scheme. Finally, the LDU performs fine inference on the reconstruction image to detect small objects. In evaluations, HODS achieved the highest mean average precision score.