SHORT TERM TRAVEL BEHAVIOR PREDICTION THROUGH GPS AND LAND USE DATA
Publication or External Link
The short-term destination prediction problem consists of capturing vehicle Global Positioning System (GPS) traces and learning from historic locations and trajectories to predict a vehicle’s destination. Drivers have predictable trip destinations that can be estimated through probabilistic modeling of past trips.
This dissertation has three main hypotheses; 1) Employing a tiered Markov model structure will permit a shorter learning period while achieving similar accuracy results, 2) The addition of derived trip purpose information will increase accuracy of the start of trip and in-route models as a whole, and 3) Similar methodologies of travel pattern inference can be used to accurately predict trip purpose and socio-economic factors.
To study these concepts, a database of GPS driving traces (120 participants for 70 days) is collected. To model the user’s trip purpose, a new data source was explored: Point of Interest (POI)/land use data. An open source land use/POI dataset is merged with the GPS dataset. The resulting database includes over 20,000 trips with travel characteristics and land use/POI data. From land use/POI data, and travel patterns, trip purpose is calculated with machine learning methods. A new model structure is developed that uses trip purpose when it is available, yet falls back on traditional spatial temporal Markov models when it is not.
The start of trip model has an overall increase of accuracy over other start of trip models of 2%. This comes quickly, needing only 30 days to reach this level of accuracy compared to nearly a year in many other models. When adding trip purpose and the start of trip model to in-route prediction methods, the accuracy of the destination prediction increases significantly: 15-30% improvement of accuracy over similar models between 0-50% of trip progression. Certain trips are predicted more accurately than others: work and home based trips average of 90% correct prediction, whereas shopping and social based trips hover around the 50% mark. In all, the greatest contribution of this dissertation is the trip purpose methodology addition and the tiered Markov model structure in gaining fast results in both the start of trip and in-route models.