Tiny machine learning (TinyML) in IoT systems exploits MCUs as edge devices for data processing. However, traditional TinyML methods can only perform inference, limited to static environments or classes. Real case scenarios usually work in dynamic environments, thus drifting the context where the original neural model is no more suitable. For this reason, pre-trained models reduce accuracy and reliability during their lifetime because the data recorded slowly becomes obsolete or new patterns appear. Continual learning strategies maintain the model up to date, with runtime fine-tuning of the parameters. This paper compares four state-of-the-art algorithms in two real applications: i) gesture recognition based on accelerometer data and ii) image classification. Our results confirm these systems' reliability and the feasibility of deploying them in tiny-memory MCUs, with a drop in the accuracy of a few percentage points with respect to the original models for unconstrained computing platforms.

Infra-red (IR) cameras have found widespread use in many different fields. The most common ones are generally related to industrial applications, particularly maintenance and inspections activities. In the domain of surveillance, instead, they are mostly used for threat detection and security purposes. Pushed by cost reduction and the availability of compact sensors, intelligent IR cameras are gaining popularity in the field of Internet-of-Things, in light of the valuable information made available by the acquired data. Unfortunately, the achievable overall quality is not always satisfactory. For example, low-resolution devices, noise, or harsh environmental conditions, like high temperatures on sunny days, can degrade the quality of the thermal images. This paper presents the development of a portable, low-cost, and low-power thermal scanner prototype consisting of a thermal sensor assisted by a grayscale camera. The prototype is completely made using COTS components and provides 80 × 60 IR and 160 × 120 grayscale images, mostly used to collect and validate the IR-based data. Our application focuses on people detection, for which we present a suitable learning framework together with the corresponding IR dataset, collected and annotated via the paired grayscale images.