Improving classification accuracy of acoustic real-world urban data using sensors physical redundancy

Latest advances in modern society together with the increase of the population living in urban areas have transformed these environments into noisy spaces. Current regulations limit the amount of noise-per-source that can impact the population. Hence, automatically identifying acoustic events in urban environments is of great interest for public administrations to preserve citizens' health. Therefore, alternatives that are typically composed of expensive sensing devices committed to individually survey a specific area have been researched. The purpose of this paper is to assess the performance of an alternative approach composed of a low-cost acoustic wireless sensor network that takes advantage of physical redundancy. Specifically, the evaluated system incorporates a deep neural network running in each sensor node and a distributed consensus protocol that implements a set of heuristics to benefit from the classification results of neighboring nodes surveying the same area (i.e., physical redundancy). To evaluate this system, real-world acoustic data were collected simultaneously from four different spots of the same crossroad in the centre of Barcelona and further processed by the system. Obtained results suggest that physical redundancy of sensors improves the classifier's confidence and increases the classification accuracy.