Spatio-temporal Network Analysis for Detecting Traffic Accident Clusters

Transcript

1. Spatio-temporal Network Analysis for Detecting Traffic Accident Clusters 〇 Keisuke ANDO1 Yusuke KUNIYOSHI1 Kazuhiko SHIMIZU1 Takeshi UCHITANE1 Naoto MUKAI2 Kazunori IWATA3 Nobuhiro ITO1 Yong JIANG3 1Aichi Institute of Technology 2Sugiyama Jogakuen University 3Aichi University IIAI AAI 2023 14th International Congress on Advanced Applied Informatics
2. Objective of the Study 2 Detecting hotspots where traffic accidents occur frequently Traffic accident clusters Traffic accident point Road Cannot detect where locations are danger Traffic accident cluster
3. Previous Studies 3 Studies about detecting traffic accident hotspots Estimate the risk on the spatio-temporal network Detect traffic accident hotspots on the network NOT considering temporal effects NOT detecting the hotspots (Xie et. al., 2013) (Romano and Jiang, 2017) hotspots road time low-risk high-risk
4. Key Points of the Study 4 Extending the method of Xie et. al. temporal direction Traffic accident hotspots on the network Traffic accident clusters (hotspots) on the spatio-temporal network Can detect where and when are danger
5. Proposed Method 5 Check the spatial properties of traffic accidents have not changed over time in the study area Estimate the probability of the occurrence of traffic accidents using kernel density estimation Detect locations and times where the probability of the occurrence is significantly higher 1 2 3 Spatial properties remained consistent 3 Steps to detect traffic accident clusters
6. 6 General method for studying the distribution of spatial events Estimating the probability of an event from its density based on the location of the event Estimating the probability density of an event from its spatial density based on the event points Network KDE is suitable for analyzing traffic accidents Planar KDE Network KDE Accident points KDE Kernel Density Estimation
7. STNKDE 7 Temporal extension of Network KDE , location time Spatio-Temporal Network Kernel Density Estimation Impact of the accident on the network direction Impact of the accident on the temporal direction Point of the accident Impacts of the accident
8. Proposed Kernel Functions 8 Kernel functions on the network and temporal direction Temporal direction Network direction 𝑓 , 𝑥, 𝑡 =
9. Difference from Previous Kernel Functions 9 Differences in estimation at intersections Previous kernel functions Proposed kernel functions Considering only path distance Considering path distance and number of branches Overestimation at intersection Appropriately estimate the risk of accidents at intersections
10. Procedure of Detecting the Clusters 10 Hypothesis testing using Monte Carlo method (Randomization Tests) Accident location and time Traffic accident cluster Detect Monte Carlo simulation 1 Applying STNKDE 2 Hypothesis Testing 3
11. Discretized unit the spatio- temporal network Step 1: Monte Carlo Simulation 11 Generate pseudo distribution of traffic accidents 1 incident 4 incidents 5 incidents 0 incidents shuffle arixels 1 2 𝑛 arixel 1 h 20 m
12. Step 2: Applying STNKDE 12 Estimation of the probability density functions from each data Probability density function of actual accidents Probability density function of pseudo accidents data 1 2 𝑛
13. Step 3: Hypothesis Testing 13 Extracting arixels with significantly higher probability density Extract Compare (Test) Traffic accident cluster
14. Ways of shuffling 14 Two types simulations based on different hypotheses Shuffles all arixels Shuffles arixels with at least one traffic accident Complete random Conditional permutation This is just one of the points where traffic accidents have happened before This is just one of the points where traffic accidents have happened before This is just one of the points where traffic accidents have happened randomly This is just one of the points where traffic accidents have happened randomly Difficult to visualize ❌ too many significant arixels
15. Target Data 15 Place with the most accidents in Aichi Prefecture, Japan 4.5 km 6.5 km Study area1 2015 – 2020 Study period Sakae Station Nagoya Station 1 https://www.openstreetmap.org/ 1,101,000 Number of arixels (Network size) 15,297 Number of accidents
16. Experiment Settings 16 Comparing clusters detected by different bandwidth Bandwidth (network, time) Spatio-temporal sphere of impact centered on the location and time of traffic accident occurrence accident point bandwidth (250m, 2.5h) (100m, 2.5h) wide bandwidth narrow bandwidth Which are good parameters?
17. Results of Experiment 17 large, cohesive traffic accident clusters can be visualized Visualization of traffic accident clusters from the experiment road network 24 0 Time 6 12 18
18. Evaluation of Proposed Kernel Functions 18 Proposed kernel function Previous kernel function (Romano and Jiang) Comparison of estimated probability densities near the intersection Proposed kernel function does not have a high probability density Proposed kernel function is more reliable than previous
19. Which are good parameters? 19 Bandwidth: (100m, 2.5h) Wide bandwidth vs. Narrow bandwidth Wide bandwidth is better parameters Clusters are NOT cohesive Bandwidth: (250m, 2.5h) Clusters are cohesive
20. • We proposed a method for detecting hotspots where traffic accidents occur frequently. • The proposed method is a natural extension of previous method. • Results of experiment show the proposed method can detect traffic accident clusters. 20 Summary Spatio-temporal network analysis for detecting traffic accident clusters