Resumen
As the COVID-19 pandemic emerged, everyone?s attention was brought to the topic of the health and safety of the entire human population. It has been proven that wearing a face mask can help limit the spread of the virus. Despite the enormous efforts of people around the world, there still exists a group of people that wear face masks incorrectly. In order to provide the best level of safety for everyone, face masks must be worn correctly, especially indoors, for example, in shops, cinemas and theaters. As security guards can only handle a limited area of the frequently visited objects, intelligent sensors can be used. In order to mount them on the shelves in the shops or near the cinema cash register queues, they need to be capable of battery operation. This restricts the sensor to be as energy-efficient as possible, in order to prolong the battery life of such devices. The cost is also a factor, as cheaper devices will result in higher accessibility. An interesting and quite novel approach that can answer all these challenges is a TinyML system, that can be defined as a combination of two concepts: Machine Learning (ML) and Internet of Things (IoT). The TinyML approach enables the usage of ML algorithms on boards equipped with low-cost, low-power microcontrollers without sacrificing the classifier quality. The main goal of this paper is to propose a battery-operated TinyML system that can be used for verification whether the face mask is worn properly. To this end, we carefully analyze several ML approaches to find the best method for the considered task. After detailed analysis of computation and memory complexity as well as after some preliminary experiments, we propose to apply the K-means algorithm with carefully designed filters and a sliding window technique, since this method provides high accuracy with the required energy-efficiency for the considered classification problem related to verification of using the face mask. The STM32F411 chip is selected as the best microcontroller for the considered task. Next, we perform wide experiments to verify the proposed ML framework implemented in the selected hardware platform. The obtained results show that the developed ML-system offers satisfactory performance in terms of high accuracy and lower power consumption. It should be underlined that the low-power aspect makes it possible to install the proposed system in places without the access to power, as well as reducing the carbon footprint of AI-focused industry which is not negligible. Our proposed TinyML system solution is able to deliver very high-quality metric values with accuracy, True Positive Ratio (TPR), True Negative Ratio (TNR), precision and recall being over 96% for masked face classification while being able to reach up to 145 days of uptime using a typical 18650 battery with capacity of 2500 mAh and nominal voltage of 3.7 V. The results are obtained using a STM32F411 microcontroller with 100 MHz ARM Cortex M4, which proves that execution of complex computer vision tasks is possible on such low-power devices. It should be noted that the STM32F411 microcontroller draws only 33 mW during operation.