Options
2016
Conference Paper
Titel
Deep person re-identification in aerial images
Abstract
Person re-identification is the problem of matching multiple occurrences of a person in large amounts of image or video data. In this work we propose an approach specically tailored to re-identify people across different camera views in aerial video recordings. New challenges that arise in aerial data include unusual and more varied view angles, a moving camera and potentially large changes in environment and other inuences between recordings (i.e. between ights). Our approach addresses these new challenges. Due to their recent successes, we apply deep learning to automatically learn features for person re-identification on a number of public datasets. We evaluate these features on aerial data and propose a method to automatically select suitable pretrained features without requiring person id labels on the aerial data. We further show that tailored data augmentation methods are well suited to better cope with the larger variety in view angles. Finally, we combine our model with a metric learning approach to allow for interactive improvement of re-identification results through user feedback. We evaluate the approach on our own video dataset which contains 12 persons recorded from a UAV.