COVID-19 pandemic has exposed the vulnerability inherent in our health care ecosystem and called for innovative systems that are portable and scalable to enable rapid deployment. This necessity has opened the door for thermal imaging systems to become integral components of the clinical infrastructure. We all might have seen the crucial role of thermal camera in airports, malls, hospitals and other public places. This adaptation was not just a response to the immediate challenges posed by COVID-19 but also a forward-looking investment in technologies that can fortify our healthcare infrastructure for future crises.
In this blog, we will delve into the diverse applications where thermal imaging has proven invaluable to mitigate the spread of COVID-19.
Thermal Imaging for Fever Screening
As the world grappled with the rapid spread of COVID-19, temperature monitoring swiftly became a standardized protocol in public places. Elevated body temperature, a common symptom of the virus, prompted the need for rigorous screening measures to identify potential carriers and mitigate the risk of transmission. Traditional handheld forehead thermometers played a crucial role in this initial phase, but their limitations soon became apparent.
While handheld thermometers were effective, they posed logistical challenges. Their operation required close proximity between the personnel conducting the screening and the individuals being screened—within a mere 3 feet. This proximity heightened the risk of infection for both parties involved. In crowded public spaces, where the need for efficient and safe screening is paramount, the limitations of traditional methods became evident.
In response to these shortcomings, thermal imaging cameras emerged as a transformative solution. Thermal cameras completely automated the fever screening process and allowed for non-contact temperature monitoring from a safe distance. This not only enhances efficiency but also significantly reduces the risk of potential virus transmission during the screening process.
The automation of fever screening through thermal cameras involved two steps:
(i) Automated Face Detection: Human face comprises an extensive network of blood vessels, and blood flow is closely regulated to maintain thermal equilibrium. Changes in core body temperature are often reflected in the thermal patterns of the face due to alterations in blood flow. Additionally, facial tissues are thin, enabling a direct transfer of heat and making the face sensitive to temperature changes.
Therefore, the first step in automation of fever screening involves the identification and detection of bounding boxes corresponding to potential face regions in the images. To achieve this, widely used face detector deep learning networks such as MobileNet-SSD, YOLO, among others, were popularly employed.
(iii) Temperature Measurement: Once the face region is detected, the subsequent challenge involves the identification of body temperatures from the facial temperatures. Various heuristics were employed to determine which facial region to consider for approximating body temperature. This included measuring the temperatures on inner canthus region of the eye, forehead, nose, ears, mouth (open and closed), cheeks among other potential indicators.
A recent exploration [1] has demonstrated that full-face maximum temperatures provide the best performance followed closely by the maximum of a wider inner canthi region in approximating the body temperature.
Thermal Imaging for ensuring adherence to social distancing and mask compliance
Thermal imaging cameras were also employed for implementing social distancing and also flagging the people who are not complying with mask guidelines amid the pandemic. You might wonder why one can not use standard visual cameras for these tasks. The answer lies in the inherent limitations of visual cameras, which necessitate a minimum illumination for effective operation. For example, you cannot deploy the visual cameras at night time. On other hand, thermal imaging does not require any minimum lighting constraints.
In the literature, a comprehensive analysis comparing the performance of visual and thermal-based imaging in diverse lighting conditions highlights the advantages of thermal imaging. This is evident from the below table that compared the performance of visual and thermal based mask classifiers under different lighting conditions. Visual cameras performance was better in the presence of high illumination, but it experienced degradation in performance as illumination decreases. On the other hand, thermal imaging resulted in consistent performance despite the variations in illumination. This disparity emphasizes the reliability of thermal imaging in scenarios where lighting is suboptimal or variable, making it an invaluable asset for maintaining adherence to social distancing and mask compliance.
Lighting Conditions | YOLOv3 |
Visual trained (%) | Thermal trained (%) |
0-25 lux | 54.60 | 100 |
25-75 lux | 94.68 | 100 |
75-150 lux | 98.91 | 100 |
>150 lux | 98.72 | 100 |
Table I: Performance of Mask classifier trained with visual images and thermal images under different lighting conditions. [2]
Beyond its technical advantages, the use of thermal imaging addresses crucial privacy and ethical considerations associated with surveillance. By design, thermal imaging prioritizes privacy as it solely captures temperature measurements without recording actual faces. This innovative approach ensures the protection of participants' privacy, fostering a balance between effective monitoring and ethical considerations in the pursuit of public health and safety.
Finally, this surveillance can be combined with the above fever screening module to create a holistic system that performs fever screening, mask classification and distance monitoring. One such example of the end-to-end system is highlighted below. For the purpose of automated surveillance for mask classification and distance monitoring, deep learning modules such as YOLO were typically employed.
Open Source Datasets
If you want to explore more on the above problems, below are open source datasets:
NTIC COVID-19 Screening Dataset
NTIC COVID-19 Screening dataset is an open source dataset from Niramai Health Analytix. This dataset was collected on Indian population during COVID-19. Broadly, This comprehensive dataset is a consortium of multiple datasets for different COVID-19 screening purposes.
(a) Thermal Surveillance Dataset: Thermal Surveillance Dataset comprises a rich collection of thermal images capturing individuals who walked-in into public premises such as offices, malls, and railway stations. It comprises a total of 902 thermal images, where each image within this dataset may feature more than one individual, providing a realistic representation of crowded public spaces. A total of 1,567 individuals are part of this dataset.
(b) Augmented Surveillance Dataset: This dataset comprises thermal-like images generated from visual RGB images using data transformations. Overall, this dataset comprises 543 images with 434 people wearing masks and 109 people without masks.
(c) Lighting Dataset: This is a controlled dataset that captured both RGB and Thermal images at same time and only one person at a time under 4 different lighting conditions. Overall, 420 thermal images from 25 participants under 4 different lighting conditions were part of this dataset. For each of these thermal images, corresponding visual images are also captured.
All the above datasets comprise the bounding box regions corresponding to individual faces for all the individuals along with whether each person is wearing masks or not. This allows one to develop ML algorithms for below problem statements
- Identification of face bounding boxes from thermal images: This involves detecting bounding boxes corresponding to face regions of individuals in a thermal image.
- Detection of Mask from thermal images: This involves identifying whether each of the individual in a given thermal image is wearing mask or not.
- Synthesis of thermal-like images given a visual image: Leveraging this thermal dataset and the other RGB datasets, it would be interesting to synthesize thermal-like images from a visual image. This data synthesis would help in creating larger thermal datasets.
[2] Zhou Y, Ghassemi P, Chen M, McBride D, Casamento JP, Pfefer TJ, Wang Q. Clinical evaluation of fever-screening thermography: impact of consensus guidelines and facial measurement location. Journal of Biomedical optics. 2020 Sep;25(9):097002.
Comments
Post a Comment