Anomalous vehicular traffic detection through spectral techniques and semi&un-supervised learning models
Export citation
Abstract
The early development of the Internet of Things has allowed the construction of collaborative systems capable of responding effectively to events captured by sensors and devices. Also, it has given the ability to share information among themselves. This paradigm has opened up the development of initiatives such as Smart Cities.
The main objectives of a Smart City are to improve the standards of people's daily lives and to deal with the various urban problems to satisfy the needs of present and future generations; one of them is related to mobility. For example, poor transport management negatively impacts a city, as there is an increase of air and noise pollution, also in trip times for drivers, energy consumption and vehicular congestion.
Among the range of problems related to mobility, we find the anomalous vehicular traffic. This problem can be caused by different reasons, for example, an accident, an event, road works, or a natural disaster. When performing the detection of this problem, it is possible to take short-term and long-term decisions, for example, by alerting drivers of the anomaly and allow them to make better decisions during their journey.
Previous researches on the detection of AVT bases their solutions on using inductive loop sensors, video surveillance and crowdsourcing systems. However, these solutions are limited in the sense that they focus mainly on creating new algorithms, and do not pay attention to the underlying information that can be extracted from vehicular traffic.
In this work, vehicular traffic is classified as a univariate time series where only there is a feature available. Therefore, feature extraction is relevant when information is scarce, since by having new features there is a significant contribution of information to enhance anomaly detection models, by facilitating their learning process and improving its performance.
Ramp Metering is used at freeway on-ramps and regulates the vehicular traffic when entering to freeways according to current traffic conditions. With the use of this device, the improvement in vehicular congestion during rush hours has been demonstrated. In this time, the device is active. However, the problem arises when there is anomalous vehicular traffic in hours where Ramp Metering is deactivated and a good manage on the vehicular traffic is no longer possible.
To address the lack of features by using univariate time series and the problem of using the Ramp Metering in extended hours, we propose a methodology for feature extraction to enhance anomaly models to detect anomalous vehicular traffic at freeway on-ramps. As a first step, an algorithm for missing values imputation is proposed, followed by temporary, spectral, and aggregates features extraction, all of them in different time aggregations. To finally use unsupervised models and perform the anomaly detection in semi-supervised and unsupervised learning. The unsupervised models used in this work are: Isolation Forest, Local Outlier Factor, One-Class Support Vector Machine and Angle-Based Outlier Detection.
The methodology was evaluated in a real vehicular traffic and synthetic databases. Experimental results show that the spectral, temporary and aggregation features enhance the detection of anomalous vehicular traffic. The Isolation Forest algorithm is compared with a literature's algorithm based on Markov-modulated Poisson processes, and obtained the best performance.
Collections
The following license files are associated with this item: