by Adel Bolbol
Bolbol, A Cheng, T Skarlatidou, A and Tsapakis, I, 2011. A GIS-Machine Learning Approach for Inferring Travel Mode From Sparse GPS Data. Proceedings of 9th International Conference on Survey Methods in Transport (ISCTSC), Termas Puyehue, Chile, November 14-18, 2011. (Accepted)
Travelling is a very interesting yet complex activity that happens on a daily basis. Travel behaviour has proven to be... more
Travelling is a very interesting yet complex activity that happens on a daily basis. Travel behaviour has proven to be very challenging to model due to its complexity, diversity and size of the data. Understanding this kind of activity is important for applications such as measuring time expenditures, quality of life, tourist activity, strikes’ impact on transportation and other environmental impacts. In order to understand that kind of behaviour; research has focused on using machine learning methods to infer meaningful information about people’s trips. Among these types of information is the type of travel mode. This could include walking, cycling, cars, trains … etc. Studies attempt to do this inference from positional data obtained from sensor devices such as GPS. This inference could largely replace or complete a lot of the feedback required by users when labelling and tagging their travel diaries. This helps overcoming low response rates, which is one of the major problems of travel diaries, by taking the burden off the user to complete their recall surveys and other user-related privacy issues.
Previous literature attempting to identify transport modes for multi-modal trajectories holds numerous disadvantages. These include the negligence of the network matching process, ignoring intermediary modes such as the “stationary” mode, the lack of details of the accuracy measurement procedure, the lack of research on map matching for multi-modal trajectories, relying on likelihood and much more. Therefore, the challenge is to produce a model that overcomes all these disadvantages and that is capable of handling the diverse and complex multi-modal nature of GPS data.
In our previous work, we tested for the best configuration for GPS data collection to attain an appropriate quality level for inferring the travel mode and correct trip route purpose. We tested first for the best epoch rate collection appropriate for that purpose. We performed that by testing some user trips with respect to positional error, distance covered, speed and origin-destination accuracies. Results revealed that the best epoch rate to be used is around 60 seconds. We also tested for the appropriate ways of carrying around GPS devices for the same purpose. Results revealed that different devices perform differently. However, for the devices we used in our tests we did not note significant differences for the method of device handling. However also, it was recommended not to carry devices in lower pockets (like in jeans pockets) to insure good coverage.
In other previous work, we introduced Geotraveldiary.com as an online user-generated travel diary system that records users’ trails based on Web GIS technology. The online server-side application allows the users to visualize their trails and manually edit, move or remove points from their tracks on a map-based interface. This in turn helped increasing the size of our database with positional data labelled with travel information for a large city (such as London in this case). As a result, we obtained user-generated information from this system to learn and infer travel behaviour indicators such as travel mode and the purpose of the trip. And the more data was added and labelled, the more training data ready to be studied and used.
In this work we attempt to build up on our previous work by identifying travel modes solely from sparse GPS data. We perform this inference by creating a framework consisting of a combination of machine learning and GIS methods to identify different modes. These include Support Vector Classification (SVC), GIS analysis, transition probability and fuzzy logic. The framework consists of a sequence of steps, each containing one of these methods. Most steps act as a verification of results from previous ones. Some also work iteratively with others to enhance the accuracy of both the segmentation and classification. The application area is the city of London due to its complexity and the diversity of its transportation networks.
The framework introduced in this work classifies a “stationary” mode as an intermediary mode in the inference model. This is an important travel mode type since it occurs with high frequency in any GPS track. It also usually occurs as a transitional mode between two different modes such as waiting for a bus at a bus stop at the timeline between a “walk” and a “bus” mode. It also occurs as instances within some mode segments such as “walk”. Taking this transition probability into account increases the accuracy of labelling; and hence avoiding confusion within the inference model. This has been usually ignored in similar previous research, which our method attempts to overcome.
Another advantage of this method over previous studies is the usage of different transport networks to match GPS points to. Among these networks are the bus, tube, train, road, and pathway networks. The advantage of this is eliminating most of the GPS error, which increases the accuracy of speed data used for classification. Matching each segment of data to each network type also verifies the results of the classification obtained from other methods (if not does the job entirely) by assigning the appropriate modes according to the matched network.
We provided a novel approach to performing location-based activity recognition. In contrast to existing techniques, our approach uses one consistent framework for both inferring travel mode and network matching. An initial segmentation of the track first takes place to break the track into different potential segments each of a specific travel mode. The framework then identifies the “stationary” mode by applying a novel spatial clustering algorithm we developed for that purpose. The “tube” and “train” modes are made identifiable using a sequence of network matching and fuzzy logic algorithms. The “car”, “cycle” and “walk” modes are differentiated using an iterative approach of a group of methods namely; SVC, network matching and GIS analysis on a local feature database. A modal transition matrix is applied afterwards to verify the classification and to apply merges to any misclassified trip segments.
Results reveal that a good classification of most travel modes is attained with an overall accuracy above 85%. The tube and train modes were easily identifiable by our algorithm with an accuracy of around 91%. Also the “car”, “cycle” and “walk” results reveal an increase in classification accuracy compared to previous studies. That, as well as a significantly superb accuracy of above 95% for the “stationary” mode; which helped increase the accuracy of the segmentation process and hence the classification as a whole. As a result, the framework demonstrates a great ability to make use of geographical databases and GPS data to infer travel modes and hence attempting to overcome some of the major disadvantages of travel diaries.
Download (.pdf) 
