Thermal Nostril Tracking Using YOLOv4-Tiny

The respiration rate (RR) is an essential parameter to ascertain human health. However, the existing methods of RR measurement are contact-based and have many drawbacks. In the proposed work, the breathing signals (BS) are acquired by plotting the variation in the temperature across the nostrils using infrared thermography (IRT). The acquired BS are filtered using a digital Savitzky-Golay (SG) filter, and the impact of the filter’s frame length on the signal’s SNR is evaluated. Further, ‘You Only Look Once (YOLOv4-tiny)’, a deep-learning-based, and extremely fast object detector is implemented to automate nostril tracking in real-time. Performance of YOLOv4-tiny is analyzed using the metrics like the average intersection-over-union (IoU), mean average precision (mAP), precision, recall, and the F1-score, at IoU thresholds of 0.5 and 0.75 each. For an IoU threshold of 0.5, the optimal model is observed to have 0.74 average IoU, 0.978 mAP, with precision, recall, and F1-score of 0.98 each, at the 2000^th training iteration. While, for an IoU threshold of 0.75, the following optimal metrics are recorded at the 6000^th training iteration–the average IoU of 0.66, mAP of 0.72, and precision, recall, and the F1-score of 0.8 each. The model generated nostril detections with an average time of 3.63 millisecond.