Traffic Level Monitoring in Urban Scenarios with Virtual Sensing Techniques Enabled by Embedded Machine Learning

Urban traffic is one of the most significant causes of pollution, significantly contributing to the level of gaseous emissions and to high noise levels in city centres. Noise can then be assumed as an indicator to estimate traffic levels and the overall air quality. The scope of this paper is to demonstrate the chance of estimating traffic levels from the acquisition of audio samples, exploiting embedded machine learning techniques. To this aim, a set of embedded devices is used to collect an ad-hoc dataset for the training phase and then to implement the algorithm for the traffic level extraction. The prototype of the sensor node used for traffic estimation is based on a microcontroller featuring limited performances and reduced power consumption, allowing to set up an autonomous platform in charge of locally predicting the values without the need to transmit the sampled audio files to a remote data processing unit. Tests demonstrate that the embedded machine learning algorithm is able to correctly detect the traffic level around 95% of times, even with a complexity reduction of the algorithm itself, suggesting the possible implementation of a pervasive monitoring infrastructure based on a number of autonomous virtual sensors.