Traffic Load Estimation (TLE) is increasingly adopted in public road infrastructures to regulate the access and limit heavy vehicles circulation. Standard approaches to TLE are based either on installing dedicated sensors such as intelligent cameras or infrared sensors or using existing smartphone sensors. However, both approaches have severe limitations, as often dedicated sensors are power-hungry and expensive to install and maintain, whereas smartphone-based approaches critically rely massively on users collaboration. More recently, researchers have started investigating TLE approaches using networks of accelerometers that are often already installed on critical road elements such as viaducts and bridges for Structural Health Monitoring (SHM) purposes. Specifically, in previous solutions, the detection and counting of vehicles was based on unsupervised anomaly detection and did not use any labeled data. While this simplifies the system's setup, it also makes full validation impossible.In this work, we investigate the TLE problem using a supervised learning approach for SHM-sensor-based TLE for the first time. In particular, we use a relatively short recording session from a smart camera to label acceleration data with the corresponding number (and type) of passing vehicles. Labeled data are then fed to a Machine Learning (ML) model trained as a regressor to estimate the vehicle count corresponding to each input sample. We perform an extensive comparison among different types of ML models, both classic and deep. Our experiments find that the highest accuracy is achieved by a Support Vector Regressor (SVR) combined with simple feature extraction, which can reach a Mean Absolute Error (MAE) of 0.47 light vehicles and 0.21 heavy vehicles. This corresponds to a 9.8x and 8.1x error reduction compared to previous unsupervised solutions, respectively. Lastly, we show that our approach lends itself to an energy-efficient implementation on a real SHM gateway.
Alessio Burrello, Giovanni Zara, Luca Benini, Davide Brunelli, Enrico Macii, Massimo Poncino, et al. (2022). Traffic Load Estimation from Structural Health Monitoring sensors using supervised learning. SUSTAINABLE COMPUTING, 35, 100704-100719 [10.1016/j.suscom.2022.100704].
Traffic Load Estimation from Structural Health Monitoring sensors using supervised learning
Alessio Burrello
;Luca Benini;Davide Brunelli;
2022
Abstract
Traffic Load Estimation (TLE) is increasingly adopted in public road infrastructures to regulate the access and limit heavy vehicles circulation. Standard approaches to TLE are based either on installing dedicated sensors such as intelligent cameras or infrared sensors or using existing smartphone sensors. However, both approaches have severe limitations, as often dedicated sensors are power-hungry and expensive to install and maintain, whereas smartphone-based approaches critically rely massively on users collaboration. More recently, researchers have started investigating TLE approaches using networks of accelerometers that are often already installed on critical road elements such as viaducts and bridges for Structural Health Monitoring (SHM) purposes. Specifically, in previous solutions, the detection and counting of vehicles was based on unsupervised anomaly detection and did not use any labeled data. While this simplifies the system's setup, it also makes full validation impossible.In this work, we investigate the TLE problem using a supervised learning approach for SHM-sensor-based TLE for the first time. In particular, we use a relatively short recording session from a smart camera to label acceleration data with the corresponding number (and type) of passing vehicles. Labeled data are then fed to a Machine Learning (ML) model trained as a regressor to estimate the vehicle count corresponding to each input sample. We perform an extensive comparison among different types of ML models, both classic and deep. Our experiments find that the highest accuracy is achieved by a Support Vector Regressor (SVR) combined with simple feature extraction, which can reach a Mean Absolute Error (MAE) of 0.47 light vehicles and 0.21 heavy vehicles. This corresponds to a 9.8x and 8.1x error reduction compared to previous unsupervised solutions, respectively. Lastly, we show that our approach lends itself to an energy-efficient implementation on a real SHM gateway.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
