Classification of Travel Data with Multiple Sensor Information using Random Forest

Recently, a lot of studies have been focused on the use of smartphones for automatic detection of transportation mode. This task is made easy by the availability of sensors like accelerometer and GPS in modern smartphones. The advantages include the increased accuracy which was partially lost due to underreporting in case of conventional travel surveys. In this paper Probe Person data collected by 46 participants in three different cities of Japan, namely Niigata, Gifu and Matsuyama, was used. Although the data, comprising of acceleration and GPS information, was collected by a wearable device but the same can be achieved very easily with the help of smartphones. In order to address the most important problem of continuously changing position of smartphone during the trip, only resultant acceleration was taken. In addition, personal characteristics like age and gender were also included. Regarding GIS information, distance and time calculated by Google Maps for both driving and walking was introduced to increase the prediction accuracy. Random Forest was applied for the purpose of prediction. 70% of the data was randomly selected to train the algorithm and rest 30% was used to test it. Prediction was done among four different modes; walk, bicycle, car and train. The results are quite promising with an overall prediction accuracy of more than 99.6% for all three cities. A slight improvement in the prediction accuracy is achieved by selecting the best features for the classification purpose.