Route-Aware Edge Bundling for Visualizing Origin-Destination Trails in Urban Traffic Wei Zeng 1 , Qiaomu Shen 2 , Yuzhe Jiang 2 , Alex Telea 3 1. Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences 2. The Hong Kong University of Science and Technology 3. University of Groningen
Contents § Introduction o OD Trails in Urban Traffic o Prior Edge Bundling Methods o Limitations of KDEEB § Route-Aware Edge Bundling o Preprocessing: Ø map matching → hierarchical route structure construction → trail abstraction o Bundling Ø optimal kernel size setting → density map generation o Evaluation Ø Bundle termination Ø Bundle deviation § Conclusion and Future Work
Contents § Introduction o OD Trails in Urban Traffic o Prior Edge Bundling Methods o Limitations of KDEEB § Route-Aware Edge Bundling o Preprocessing: Ø map matching → hierarchical route structure construction → trail abstraction o Bundling Ø optimal kernel size setting → density map generation o Evaluation Ø Bundle termination Ø Bundle deviation § Conclusion and Future Work
OD Trails in Urban Traffic [Ferreira et al. 2013] § Urban traffic data, e.g., o Taxi trips in New York, Beijing, Shenzhen o Public transportation data in Singapore o Electric scooter tracks in Stuttgart § Origin-destination (OD) is a fundamental concept in transportation, summarizing (people/vehicle/good) movements across geographic locations. § Properties of OD trails in urban traffic o Locations o Times o Road network o Multi-modes [Krüger et al., 2013]
OD Trail Visualization § Density Map o Summarize trajectories and overview distribution. § Spatial Aggregation [Scheepens et al., 2011] o Partition underlying territory into appropriate areas. § Map Matching o Align position records with road network data. § Direct depiction o Directly plot trajectories into 2D/3D displays. [Andrienko and [Wood et al., 2010] Andrienko, 2011] [Krüger et al. 2018] [Kwan, 2000] [Kapler and Wright, 2004]
Prior Edge Bundling Methods § Geometry-based methods: Use control mesh to specify how similar edges are routed. Pros: Flexible to make control mesh o Cons: Constructing control mesh can be (very) slow o § Force-based methods: Model interaction between spatially close trails as a force field. Pros: No need to make external control mesh o Cons: Slow – cannot handle a few thousands trails at interactive rates o § Image-based methods: Employ image-processing methods to accelerate the bundling process. Pros: Feasible for GPU implementation – can process millions of trials at interactive o rates. Cons: No consideration of spatial constraints when applied to OD trails. o
Prior Edge Bundling Methods § Constrained Bundling : Specific constraints are considered. Ambiguity o 3D curved surfaces o Directions o Obstacles avoidance o Vector map o Vector map for Swiss commuter data [Thöny & Pajarola, 2015] Map matching Vector map
Kernel Density Estimation Edge Bundling (KDEEB) § We chose KDEEB for the basis of our method: Fast in speed, meanwhile simple enough to implement o Be able to incorporate specific constraints o n times § KDEEB pipeline Sampling o Gradient estimation o Sampling Smoothing Advection o Smoothing o Gradient Edge Estimation Advection § Iterate n times until stable layout Predefined 10 or 15 times o Automatically determined at runtime? o
Limitations of KDEEB: What is a suitable pr? § KDEEB: 5% of graph drawing size 1440 ' + 720 ' = 80.5 5% × o 1440 px P r = 40 P r = 20 P r = 80 720 px P r = 120
Limitations of KDEEB: Road neglect
Limitations of KDEEB Artifacts KDEEB ( p r = 60) KDEEB ( p r = 21) Map Matching
Contents § Introduction o OD Trails in Urban Traffic o Prior Edge Bundling Methods o Limitations of KDEEB § Route-Aware Edge Bundling o Preprocessing: Ø map matching → hierarchical route structure construction → trail abstraction o Bundling Ø optimal kernel size setting → density map generation o Evaluation Ø Bundle termination Ø Bundle deviation § Conclusion and Future Work
Route-Aware Edge Bundling § RAEB pipeline: 1) Preprocessing, 2) Bundling, and 3) Evaluation Preprocessing Bundling Evaluation Previous Mutual Density Simplified Gradient Hierarchical Raw Road Image Information kernel Map Route Structure Map Road Network gradient Network size estimate Frechet edge No Route Awareness splat Distance advect save Stop? Yes Smooth Sampled Bundled smooth Abstract Map OD Trails on Current Final Raw Urban Bundles Edges Graph Matching OD Trails Image Image Road Network Traffic Input Output Road Level of details OD trails network
Preprocessing § Build a simplified hierarchical road and traffic network representation. Map matching: shortest path for OD only, ST-matching for GPS traces o Hierarchical structure construction: route length, road hierarchy, flow magnitude o Trail abstraction: route awareness ( p ra ) o OD Trails & Hierarchical Trail Abstraction Road network route structure
Bundling § KDEEB applied to the hierarchical structure. Optimal kernel size setting o Density map generation o P r
Evaluation § Termination: Bundle stability ( p s ) to determine when to stop iteration § Bundle deviation: To determine the quality of the produced result
Contents § Introduction o OD Trails in Urban Traffic o Prior Edge Bundling Methods o Limitations of KDEEB § Route-Aware Edge Bundling o Preprocessing: Ø map matching → hierarchical route structure construction → trail abstraction o Bundling Ø optimal kernel size setting → density map generation o Evaluation Ø Bundle termination Ø Bundle deviation § Conclusion and Future Work
Application 1: Synthetic Data
Application 1: Synthetic Data
Application 2: NYC Taxi fake fake real bridges bridges bridge fake bridges
Application 2: NYC Taxi (a) Map Matching (b) KDEEB ( = 60) (c) KDEEB ( = 40) (d) KDEEB ( = 21) fake real bridges real real bridge bridge bridge (e) RAEB ( = 21, = 0) (f) RAEB ( = 21, = 1) (g) RAEB ( = 21, = 3) (h) RAEB ( = 21, = 5)
Application 3: Shenzhen Taxi Airport Beihuan Ave G4 Binhai Ave G107 (a) Map Matching (b) KDEEB (c) RAEB
Discussions § RAEB constrains trails to a given road network Route awareness ( p ra ): controls how bundles follow roads at a user-selected o hierarchy level. Kernel size ( p r ): determined by both the road network geometry and its o resolution in image space. Bundling stability ( p s ): automatically stops bundling when this similarity o exceeds a given threshold. § RAEB outperforms KDEEB on both synthetic and real OD trails Visually more realistic o Quantitively closer to ground-truth results o Comparable running time o § Limitations and future work Visual hints on bundle deformation o Incorporate directional bundling techniques o Local and adaptive parameter settings: p ra and p r o
Q & A Dr. Zeng Wei �� � Associate Researcher Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences Thank You! E-mail: wei.zeng@siat.ac.cn Web: zeng-wei.com
Recommend
More recommend
