Automated Tolling Solution with Novel Inductive Loop Detectors Using Machine Learning Techniques

Abstract

Vehicle classification has a prominent role in intelligent transportation systems (ITS) for traffic management and monitoring for various applications. One such application is the automated toll collection system. Here, the vehicles need to be accurately classified and verified, as the tolling amount depends on the vehicle type. There are several sensors that can do this under the homogeneous traffic conditions, with limited types of classes, existing in western countries. However, the traffic in countries like India is heterogeneous with the vehicle composition having wide variety of vehicles of varying static and dynamic characteristics. There is no classification solution available, which is efficient, accurate and cost effective, for such traffic. To address this, the present study develops an efficient and accurate vehicle classification system, keeping tolling applications as a case study, under heterogeneous traffic conditions. Out of the sensors used for automatic vehicle identification at toll plazas, inductive loop detectors (ILD) have high sensitivity and cost effectiveness and hence is selected in this study. This paper describes an efficient and novel methodology with time series, and supervised and unsupervised machine learning algorithms, for vehicle classification at toll plazas. The proposed methodology has four phases namely; signal denoising, signal segmentation, feature extraction and classification. In the signal denoising phase, a discrete wavelet transform (DWT) based denoising was implemented. A moving standard deviation algorithm was developed for signal segmentation to obtain vehicle signatures. The segmented signatures were processed through wavelet analysis to obtain features that can be used for training in the feature extraction phase. Supervised support vector machines (SVM) was used to build the vehicle classification algorithm. Unsupervised k-means method was also implemented for clustering vehicle signatures, for cases where the actual vehicle class labels are not available. A real-time vehicle classification system was developed and tested using the proposed methodology and was found to be accurate and ready for field implementation.