Manipal-UAV person detection dataset: A step towards benchmarking dataset and algorithms for small object detection

Abstract

Intelligent UAV video analysis has drawn the attention of many researchers due to the increasing demand for unmanned aerial vehicles (UAVs) in computer vision-related applications. Applications such as search and rescue, the military, and surveillance demand automatic detection of human targets in large-scale UAV images, which is very challenging due to the small size and inadequate feature representation of person objects. Despite the significant advancements in generic object detection tasks, the performance of the state-of-the-art small object detection algorithms falls below the satisfactory level due to the lack of a representative dataset and the limited information available for small objects. To facilitate advancements in UAV and small object detection research, we present a Manipal-UAV person detection dataset11https://github.com/Akshathakrbhat/Manipal-UAV-Person-Dataset. collected from two UAVs flying at varying altitudes, locations, and weather conditions. The dataset contains 13,462 sampled images from 33 videos having 1,53,112 person object instances. The videos are captured in an unconstrained environment with complex scenes covering small objects of varying scales, poses, illumination, and occlusion, making person detection extremely challenging on this newly created dataset. This article compares the characteristics of the Manipal-UAV dataset with the standard VisDrone and Okutama datasets having aerial view person objects. In addition, it provides baseline evaluation results of the various state-of-the-art object detection algorithms applied to the newly created Manipal-UAV Person detection dataset. The dataset is made publicly available at https://github.com/Akshathakrbhat/Manipal-UAV-Person-Dataset.