Integrating analog PIR sensing with TinyML inference for on-the-edge classification of moving objects

Abstract

Classification of moving objects plays an important role in different real-life applications, especially for security monitoring. Digital Passive InfraRed (PIR) devices are the most used solutions but fail to classify the type of moving objects. Furthermore, they rely on hardware logic which is not upgradable and evolutive. This thesis presents a Tiny machine learning (TinyML) research to classify moving objects in the surrounding environment based on reflected analog PIR wave patterns. The lack of a public dataset has driven this research to start by collecting primary analog PIR data of moving humans, dogs, goats, and windblown vegetables. Then, a TinyML classification model has been trained with the objective to deploy it on a resource-constrained embedded microprocessor for real-time classification inference. This edge-centric architecture design enables the preservation of both battery lifetime and wireless communication bandwidth of the prototyped smart analog PIR device. The pilot experimentation shows a performance accuracy of 90% which may be improved over time using reinforcement learning.