Detection of a Fallen Person from UAV Images Using Rotation Invariant Features

Abstract

In recent years, aerial photography has been used to search for victims in the event of a disaster. Searching from the sky enables quick search activities in places that are difficult to enter. In this paper we propose a method of detecting a person fallen on the ground from images taken by a camera mounted on a UAV(multicopter). Unlike pedestrians, a fallen person takes various postures, and the orientation of the head in an image is not identical. Therefore, it is necessary to develop a method which is robust to various orientations of a fallen person. In the proposed method, Ri-HOG features and Ri-LBP features invariant to object orientation are employed for representing a fallen person, and the fallen person is detected by a classifier constructed using Random Forest. The effectiveness of the proposed method was verified by experiments.