Head Tracking in Automotive Environments for Driver Monitoring Using a Low Resolution Thermal Camera

Christoph Weiss,Steffen Leonhardt,Lutz Eckstein,Marian Walter,Stefan Böshagen,Alexander Kirmas,Sören Lemcke

doi:10.3390/vehicles4010014

Abstract

The steady enhancement of driver assistance systems and the automation of driving functions are in need of advanced driver monitoring functionalities. To evaluate the driver state, several parameters must be acquired. A basic parameter is the position of the driver, which can be useful for comfort automation or medical applications. Acquiring the position through cameras can be used to provide multiple information at once. When using infrared cameras, not only the position information but also the thermal information is available. Head tracking in the infrared domain is still a challenging task. The low resolution of affordable sensors makes it especially difficult to achieve high robustness due the lack of detailed images. In this paper, we present a novel approach for robust head tracking based on template matching and optical flow. The method has been tested on various sets of subjects containing different head shapes. The evaluation does not only include the original sensor size, but also downscaled images to simulate low resolution sensors. A comparison with the ground truth is performed for X- and Y-coordinate separately for each downscaled resolution.

Highlights

Long wave infrared (LWIR) cameras are becoming more popular in the consumer section and, more suitable for monitoring applications
While object tracking or especially head tracking is a common task in image processing, the majority of the research is focused on the visible spectrum
We demonstrated an algorithm to track a face in an infrared image stream for an automotive environment

Summary

Introduction

Long wave infrared (LWIR) cameras are becoming more popular in the consumer section and, more suitable for monitoring applications. While traditional tracking algorithms are built on data from the visible domain, these algorithms might not work in the far infrared (IR) domain, because the IR image usually provides a low resolution image due to the small sensor size and less detail in texture information because of small temperature variations of a surface. Looking at the typical algorithms that work in the visual domain, Tan et al displayed the challenge of the commonly used scale invariant feature transform (SIFT) feature descriptor on LWIR images [11]. They point out the upcoming errors applying SIFT on LWIR videos showing the need of adaption of standard procedures. On an IR dataset, three trackers could achieve a high accuracy: normalized cross-correlation (NCC), discriminative correlation filter-based and discriminative classifier combined with a generative model tracker [13]

Methods

Results

Conclusion