Determining Mode of Transportation Using Mobile Phones

The public transportation system is a crucial part of the solution to the nation’s economic, energy and environmental challenges. This transportation system can be made more efficient, by studying the mobility of the users, from the sensor data collected by them. The current project have been developed to compare two methods of identifying the mode of transportation being used by a person using the location data collected from one’s cell phone. The transportation modes, such as walking, driving, etc., that a user takes, can enrich the user’s mobility with informative knowledge and provide pervasive computing systems with more context information.
The raw data is time sequenced location data retrieved from the GPS in the cell phones of the users. This data was pre-processed in order to extract relevant features like velocity, acceleration and heading change and to filter out datasets with huge gaps and redundancy. A change point segmentation algorithm based on the work by Zheng et. al. [2,3], where, the data was first broken down into segments using a loose bound of velocity and acceleration, segments smaller than a certain threshold were merged into the backward segment and consecutive segments of length smaller than another threshold were merged to form one segment was implemented using a Python Script. For every segment retrieved from the change point segmentation algorithm, the average velocity, the average acceleration and the average heading change has been calculated, and have been used as a feature for classification. The current work, uses selected trajectories of the GeoLife dataset [2,3,12], to train a decision tree based model, and uses GPS trajectory data collected from a user taking multimodal trips across Singapore, in order to test the model, using a Python script. A comparative study of the two algorithms i.e. Decision Tree and Random Forest was performed on some trajectories of the GeoLife Dataset [2] and the Random Forest algorithm was found to be more accurate than the Decision Tree algorithm with an overall accuracy of 81.2%. A comparison of the work by Reddy et al. [1] and Zheng et al. [2,3] has also been performed, where the Decision Tree and the Random Forest classifier has been applied on a point to point basis instead of a segment to segment basis. It has been found that classification after change point segmentation, gives a higher overall accuracy, and the accuracy obtained for the current dataset using a Random Forest classifier is 96.7% in discerning between motorised and non-motorised modes of transportation (walk and bus), which is higher than what has been obtained in the previous work. Therefore, the change point segmentation and classification algorithms are highly suited to build a solution to the problem of discerning between motorised and non-motorised modes of transport, using a series of time-stamped GPS locations. In the real world, this project can not only help the land transport authorities make the public transportation system more efficient, but can also help the police in law enforcement, if a system which collects real time GPS logs of people is put in place. In order to achieve these results, Weka Machine Learning Toolkit, ArcGIS, MS Excel and Scikit Learn Library of Python is being used.