Sistema de detecção de roubo com microcontrolador e deep learning

Abstract

The advancement of machine learning techniques and embedded computing has fostered the development of intelligent systems aimed at enhancing personal and vehicular security. Thus, this work presents the design, implementation, and evaluation of a prototype for automatic detection of robbery or threat situations based on real-time audio analysis. The proposed system consists of a digital microphone integrated into a single-board computer (Raspberry Pi 4), responsible for acquiring and transmitting audio signals to a remote server. On the server side, the data are processed using the Whisper speech recognition model and subsequently analyzed by a deep neural network trained to identify linguistic patterns associated with risk events. The client-server architecture employs the UDP protocol, emphasizing low latency and continuous processing of temporal windows. Experiments conducted with a dataset of 2,504 audio samples achieved an average accuracy of 96.5%, precision of 96%, and a response time below 1 s per detection window, demonstrating the technical feasibility of the proposed approach and its potential for intelligent monitoring and rapid response to security incidents.