Edge ML system for real-time footstep sound recognition

Abstract

Edge machine learning (ML) has garnered significant attention owing to the rapid proliferation of data generated by intelligent sensors and the widespread adoption of artificial intelligence (AI). The primary constraint on the application of ML algorithms pertains to the computational capabilities of sensor-equipped devices. This research project aims to construct a real-time system for sound classification, specifically targeting the recognition of footstep sounds. The HAMK Smart Research unit has commissioned this thesis. A novel algorithm has been developed for extracting temporal audio features, while the support vector machine (SVM) algorithm has been employed as the classifier. The entire signal processing pipeline is executed at the network's extreme edge. The classification results reveal the shortcomings of this approach and are subject to analysis for potential system enhancements in the future.