SonicOnEdge:A Real-Time Contextual Noise Mapping System on Edge Using GridCNN Model

Abstract

Noise maps provide city-scale perception of noise levels. However, existing mapping systems suffer from several limitations: They mostly rely on historical data, lack precise noise-source contextual information, and perform audio processing in the cloud, creating only near real-time noise maps. In this work, we explore whether audio sensing and processing can both be performed on edge devices, enabling the creation of dynamic noise maps with real-time contextual visualization. The core challenge is to design a lightweight audio classification model that runs on edge devices without compromising performance. An end-to-end prototype system has been designed for detecting vehicular honks and their subclassification in terms of vehicle type. This work designs a deep learning model, GridCNN, that takes spectrogram images of 1-sec audio samples and outperforms existing audio classification algorithms (13.23% higher accuracy) using 63.14% fewer parameters and consuming 84.21% less memory. The model deployed on a Raspberry Pi Zero 2 W edge device using TensorFlow Lite achieves 94.62% test accuracy. Suitable real-time visualization of honk information has been created on maps. For model training, suitable honk datasets (56290 samples) were prepared by collecting and annotating in-house audio data using a smartphone sensor under different traffic scenarios. Honk-related contextual details can yield valuable insights to help municipal authorities devise appropriate traffic planning policies to curb noise pollution. The proposed GridCNN model may be generalized through suitable tuning and training to enable real-time classification of other outdoor noise sources (engine noise, firecrackers, loudspeakers, etc.) as well as indoor noise classes on edge devices.