engleski

Reducing big historical location-based traffic data

With the rise of location-based services that collect user data, the collection rate and needed data storage volume increased significantly. This is especially the case in traffic systems, where users equipped with tracking devices produce large historical geospatial datasets. In this paper, we focus on the historical Floating Car Data (FCD) produced by vehicles equipped with GNSS tracking devices. The used dataset was recorded on a road network of Croatia during a five-year period between 2009 and 2014 by 4908 vehicles. This resulted in 6.55 billion records, with a storage volume of 320 GB. The used dataset can be considered as big data in terms of volume and velocity. With such large volume location-based data, the scalability issues emerged: how to reduce the traffic data without the quality loss and how to efficiently data-mine such large dataset (An et al., 2016)? The FCD is commonly used for estimating road network traffic state, especially congestion (Yang et al., 2017, Erdelić et al., 2021). The methodology used to reduce the historical dataset is divided into three steps Erdelić et al., 2021). First, the five-minute speed profiles for each road segment in the digital road map were derived from the used historical FCD dataset (Erdelić et al., 2015). In total, 561071 speed profiles were computed in five-minute time buckets with an additional 1245708 average or free-flow speed values. This resulted in the reduction of traffic data from 320 GB of raw FCD data to 428.53 MB speed profiles with used floating-point precision. To validate the aggregated data quality, the real test drive was conducted in which the real travel times were compared to the travel times estimated with speed profiles. The speed profiles produced a relative average percentage error of 4.82%. To further decrease the traffic data volume, in the second step, the observed road network of Croatia was divided into a fishnet grid. Then, for each cell, the congestion index ci ϵ [0, 1] was determined based on all of the road segments' speed profiles within the cell. The computed congestion indices were coupled with the image processing morphological closing method to determine congestion zones. Such procedure can be considered as a spatial clustering of traffic data, where the result are zones that have a significant change in the traffic pattern. Lastly, the Monte Carlo simulation of routes within the congestion zones was coupled with temporal clustering to determine Travel Time Indexes (TTIs) within the zone. This resulted in 5 to 7 TTIs per zone, ranging from -4.65% to 31.66% in value, where negative values represent shorter travel time compared to the free-flow travel time, while positive values represent an increase in the travel time, mostly in rush hours where congestion occurs. The TTIs further decreased the dataset size to 2.38 MB, without the significant loss in travel time estimation as relative travel time percentage error to real driven test routes was 4.13%. This particular has an effect in offline applications that need to store whole traffic data in the memory, which is usually limited. The TTIs show a reduction of 99.44% in data volume compared to speed profiles without the notable quality loss, as the difference in mean relative percentage travel time error to SP is 0.69%.

Floating Car Data, Speed Profiles, Congestion zones, Time-varying Travel Time Indexes

nije evidentirano