Anomalous events removal for automated traffic noise maps generation

Abstract

Road Traffic Noise (RTN) is one of the biggest pollutants in modern cities, which is known to affect public health to be the direct cause of many illnesses for their inhabitants. Until recently, RTN maps have been generated using representative static measurements collected by experts, after manually discarding all non-traffic related noise events, or Anomalous Noise Events (ANEs). However, the automation of noise measurements using Wireless Acoustic Sensor Networks (WASNs) is allowing the development of dynamic maps, which require the detection of non-traffic noise sources in real-time in order to provide accurate noise level measurements. In this work, the manual an automatic removal of ANEs are compared. The latter is based on two versions of the Anomalous Noise Event Detector (ANED) designed to detect ANEs within a WASN in real-time as a two-class classifier. The experiments on 4 h and 44 min of real-life audio data show similar error rates among all the considered annotation methods. However, the detailed analysis of the experiments reveal, on the one hand, inconsistent manual annotations in certain non-ANE labelling situations, where non-coincident expert-based decisions are observed; and, on the other hand, the decrease of the overall accuracy of the ANED-based approaches due to the large number of false alarms in the case of RTN class. Thus, although the results demonstrate the viability of the automated removal of ANEs, further research should be conducted to keep improving the automation of ANEs annotation.