Powered Two-Wheeler Detection Using Crowdsourced Smartphone Data

Using a Powered Two-Wheeler (PTW) for everyday commuting has become particularly attractive in large metropolitan areas. The industry has also picked upon the rising interest for PTW oriented services using crowdsourcing and smartphones. However, current navigation or other travel information services are not at all customized to the needs of PTWs, which are systematically being neglected from the target group. To address the needs of this diverse mode, one should first be able to identify PTWs from the rest of the vehicles. The aim of the specific study is to develop models to reveal the contributing factors which are related to the identification of PTWs using solely smartphone sensors' data and tree-based machine learning approaches. To boost the accuracy of the identification task, popular feature selection algorithms and meta-modeling strategies are implemented and evaluated. Additionally, in order to deal with the unbalanced dataset different oversampling techniques are examined. Results show that high accuracy models can be developed for the specific classification problem with the proper feature representation.