Detecting and Tracking Regions of Interest for Remote Measurement of Vital Parameters
Ladda ner
Typ
Examensarbete för masterexamen
Master's Thesis
Master's Thesis
Program
Physics (MPPHS), MSc
Publicerad
2022
Författare
Müller, Madeleine
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
The initial assessment of a mass casualty incident is essential to effectively conduct
a rescue operation. The survival rate is affected by the complexity of the incident,
and it is therefore imperative to enhance the operational capacities of emergency
medical services and civil protection agencies in mass casualty incidents. This thesis
investigates the possibilities for an unmanned aerial vehicle (UAV) to detect and
track regions of interest for remote measurement of vital parameters in visual and
thermal footage for first response triage purposes. The regions of interest are the
nose, mouth, and chest, and the UAV characteristic taken under consideration in
this thesis is image blur due to random camera motion. In this thesis, we take an object
detection approach and implement the keypoint estimation framework KAPAO
and the tracking algorithm SORT in several different experimental setups. Using
KAPAO and SORT, we achieve a good result. For the detection in the thermal
domain, the model created by transferring knowledge from the visual to the thermal
domain achieves the highest performance. We also consider adversarial training
on random motion blur, however the result shows a minimal impact on the model
performance in the presence of characteristic low-altitude UAV motion blur. Regarding
the tracking of the regions of interest, the result concludes that the SORT
algorithm improves the performance compared to assigning tracking identification
numbers based on frame-to-frame differences. The result shows that the distance
to the subjects and the image quality impacts the performance. Compared with
previous work on remote measurement of vital parameters, the algorithms of this
thesis achieve a nearly perfect score on corresponding distances. If the distances
are realizable in a UAV triage application is however unknown and has to be investigated
further. Moreover, the work of this thesis problematizes the low-altitude
UAV motion blur which poses a potential limitation in a potential UAV triage application.
An alternative could hence be to use optical stabilization measurement
for blur reduction.