GPS gives richer path data, but needs to be augmented with OD information to be useful
Machine learning is often utilized to determine the mode, while rule based methods are popular for inferring purpose
Error detection
Outliers are identified with unlikely attributes like speed over 250 km/hr
Successive filters to remove outliers: number of NSAT used to record (<3), HDOP (>5), heading and speed of 0 when GPS data trace is plotted, remove multipath error in urban canyons
Track points whose distance is less than 10m of the previous one, track points with greater than 200 km/hr speed, track points with less than 5 km/hr and time gap with previous track point of at least 1 minute, delete trips with less than 4 track points
Trip identification
The combination of dwell time, speed, and visual checks on map to determine when a set of paths is a trip
Mode detection
Input features (from GPS, GIS, Accelerometer, and Respondent's information)
duration, speed, acceleration, distance, HDOP/NSAT, heading, street network, rail station, bus routes, bus stops, ownership of vehicle
Probability Method (Fuzzy logic rules, probability matrix)
Criteria based Method
Accuracies of high 70-90% seems to be achievable
Purpose inference
Input features (From GIS, GPS, Respondents' information, other information)
Land use, POI information, duration of stay, trip ending time, frequent activity, key address, demographic data, transport mode
Rule based Method (Land-use-and-purpose-matching table, heuristic rules, closest POI matching rules, single deterministic matching method, historical data matching method)
Probabilistic method (multinomial logit model, probability calculation based on distance)