Traffic COngestion pattern classification using multi-class SVM

Abstract

Searching through and selecting data sets from large traffic databases with sensor information is often a cumbersome manual process. In this paper we present an idea that may dramatically fasten and streamline this process. The idea is to build a fast search index (COSI: COngestion Search engIne) based on meta data in combination with features from the traffic patterns along routes. Instead of ploughing through the raw detector data, COSI makes it possible to search through higher level traffic (congestion) patterns. This paper explores a first step in developing COSI: a method to classify traffic congestion patterns reconstructed with the well known adaptive smoothing method. We demonstrate the method using a preliminary set of loop detector data in the Netherlands. We extract and store individual traffic congestion patterns of these regions as images. We then use SURF (Speeded Up Robust Features) to identify the descriptors vector of each image, and apply the bag-of-features technique to generate low-dimensional representation vectors for them. Finally, we use a multiclass extended SVM algorithm to classify these patterns. The results of the method are presented and a synthesis of the findings is given. We close with some preliminary conclusions and an outlook to the further development of COSI.