Volume 8, 2020: Issue 1

 PDFDownload the article (Free)

Title:

The effect of image resolution in the human presence detection: A case study on real-world image data

Author(s):

Alexander Leipnitz, Leipzig University of Telecommunications (HfTL), Germany

Tilo Strutz, Leipzig University of Telecommunications (HfTL), Germany

Oliver Jokisch, Leipzig University of Telecommunications (HfTL), Germany

Abstract:

The automated operation of robots and flying drones is coupled to high security requirements with respect to humans and environment. Sometimes, persons have to be detected from a long distance or high altitude to allow the autonomous system an adequate and timely response. State-of-the-art Convolutional Neural Networks (CNNs) enable high object detection rates for different image data but only within their respective training, validation and test datasets. Recent studies show the limited generalization ability of CNNs for unknown data, even with merely small image changes. A typical source of such problems is the varying resolution of input images and the inevitable scaling of them to match the input-layer size of the network model. While modern cameras are able to capture high-resolution images of humans also from a longer distance, the practical input-layer sizes of networks are comparably small. Hence, we investigate the reliability of a network architecture for human detection with respect to such input-scaling effects. The popular VisDrone dataset with its varying image resolution and many relatively small depictions of humans is surveyed as well as the high-resolution AgriDrone image data from an agricultural context. Our results show that the object detection rate depends on the image scaling factor as well as on the relative size of persons. An enlarged input-layer size of the network can only partially contribute to counteract the observed effects. In addition, the detection algorithm becomes computationally more expensive by the increased effort.

Keywords:

Human detection, drone imagery, long-distance capturing, image scaling, deep learning

DOI:

https://doi.org/10.36965/OJAKM.2020.8(1)53-62

Type:

Research paper

Journal:

The Online Journal of Applied Knowledge Management (OJAKM), ISSN: 2325-4688

Publisher:

International Institute for Applied Knowledge Management (IIAKM)

Received:

28 February 2020

Revised:

4 May 2020; 7 May 2020

Accepted:

22 May 2020

Accepting Editor:

Meir Russ

Pages:

53-62