Arhitektura sistema za zaznavanje in klasifikacijo zvočnih dogodkov v 3D prostoru

Sound-related research has often received less attention compared to camera-based systems in machine perception, despite its potential for a wide range of applications, including security, environmental, and industrial monitoring. This study introduces a low-cost, high-fidelity system for real-time classification and localization of transient acoustic events. The system is capable of accurately assigning Cartesian coordinates to captured sounds, classifying event types, and associating sounds across multiple audio channels. The proposed setup employs a distributed network of 20 digital microphones and 9 Raspberry Pi boards, synchronized via Precision Time Protocol. A novel 4-microphone Angle of Arrival (AOA) sensor array is developed and spatially arranged to optimize event detection and localization. Each Raspberry Pi runs a spectral feature-based transient detector. Detected events are classified using a pre-trained deep neural network on the embedded device and localized via the Gauss-Newton non-linear least squares optimization method, using time difference of arrival (TDOA) and power ratio (RSSI) measurements. Experimental results show synchronization accuracy within 1 ms, deep acoustic event classification with up to 96% accuracy, and a root mean square error of 1.36 m for the localization accuracy. The system’s robustness is validated under various noise conditions, and its computational efficiency supports real-time deployment.