Signal Proposal
Contents
- 1 PROBLEM & MOTIVATION
- 2 OBJECTIVES
- 3 SELECTED DATASETS
- 4 DATA PREPARATION
- 5 LITERATURE REVIEW
- 5.1 1. Network-constrained spatio-temporal clustering analysis of traffic collisions in Jinghan District of Wuhan, China
- 5.2 2. Visualizing Traffic Accident Hotspots Based on Spatial-Temporal Network Kernel Density Estimation
- 5.3 3. Analysis of data on the dataset of road accidents in Paris in 2012 and 2013
- 6 APPROACH
- 7 STORYBOARD
- 8 TOOLS & TECHNOLOGIES
- 9 APPLICATION OVERVIEW
- 10 INSIGHTS
- 11 KEY CHALLENGES
- 12 TIMELINE
- 13 REFERENCES
- 14 COMMENTS
PROBLEM & MOTIVATION
Efforts by the Singapore Traffic Police in educating the public on road safety over the years have decreased the number of Fatal Accidents in Singapore by 15.7% in 2017 as compared to 2016 (Chua, 2018). Despite this improvement, accidents involving motorcyclists and elderly jaywalkers were highlighted as key concerns by the Singapore Traffic Police in 2017. This is because motorcycle accidents still accounts for more than half of the traffic accidents in 2017 and the number of elderly jaywalkers road fatalities are on the rise.
In order to attempt to reduce traffic accidents, it is important to understand where, when and who is involved in traffic accidents. A better understanding of spatio-temporal patterns specific to casualty groups aids in developing appropriate preventive measures by the authorities. In view of the benefits of applying spatio-temporal analyses to traffic accidents, our initial focus was on developing a web-based geospatial application for Singapore as there is a general lack of spatio-temporal geospatial applications developed specific to this area.
Although this serves as grounds for adapting our application to Singapore’s context, our team has used data from Leeds, United Kingdom instead, due to its easily accessible rich diversity of information. Specifically, Leeds’ traffic accident data includes coordinates of traffic accidents, time, weather and details on casualties, all of which are essential for an in-depth spatio-temporal analysis for a certain casualty group. While behavioral patterns of drivers and pedestrians in United Kingdom may not be reflective of that of Singapore’s, insights still provides some indicative directions for investigations by Singapore Land Transport Authority and Singapore Police Force.
Our application development efforts were motivated by the gap between currently available road traffic geospatial tools and analysts’ needs. Current geospatial road traffic applications mainly portray traffic accidents as point events, with little analyses generated to provide in-depth insights. As such, a web-enabled geospatial analytics tool, SIGNAL, was developed with the purpose of allowing users to conduct statistical analyses on road networks for selected target groups or environmental conditions.
OBJECTIVES
In our project, we would be creating an analytical tool for discovering network-constrained spatio-temporal patterns of traffic accidents. Specifically, it focuses on the following objectives:
- To visualise the intensity of traffic accidents on road networks cartographically on an internet-based map such as Esri
- To conduct statistical simulations on road segments to reveal evidence of clusters or correlation patterns
- To provide a user-friendly interface to for practitioners to apply relevant filters for different time periods selected
SELECTED DATASETS
The following datasets will be used for analysis, as elaborated below:
Dataset | Format | Data Attribute | Source |
---|---|---|---|
Leeds Road Traffic Accidents (2013 - 2017) | CSV |
|
UK Open Database |
Leeds Pedestrian Crossing | CSV |
|
Leeds Council Data Mill North |
Leeds School | CSV |
|
UK Open Database |
Local Authority Districts (Leeds) | SHP | UK Consumer Data Research Centre (CDRC) | |
Leeds Road | SHP | UK Consumer Data Research Centre (CDRC) | |
Leeds Motorway Junction | SHP | UK Consumer Data Research Centre (CDRC) |
DATA PREPARATION
Data Collection
Data of Leeds Traffic Accident Data (2013 to 2017), Schools and Pedestrian Crossings data were collected from the UK Open Database in CSV file format, while Leeds’ District Boundary Map, Road Network and Motorway Junctions were obtained from UK Consumer Data Research Centre (CDRC) in Shapefile (SHP) format.
Data Cleaning and Wrangling
Traffic accident point events were separated from casualty point events before removing duplicates. Unique accident points allow for visualization of intensity and spatial distribution of traffic accidents. Standardization of data, such as, ensuring columns of each CSV files and their data types are the same before reclassifying selected columns, is conducted.
Data Transformation
As our geospatial application involves the use of Network Constrained analyses, data has to be transformed to appropriate formats before relevant functions could be used. Specifically, accident traffic points have to be converted to SpatialPoints and then to Point Pattern Processes. Roads would have to be converted to SpatialLines and then to Linnet before they could be combined with Point Patterns Processes to form Linear Point Patterns. In order to constrain the Linear Point Patterns to a target area of interest, it has to be intersected with the Owin of District Boundary Map. The below figure summarises the key data transformation that takes place.
LITERATURE REVIEW
1. Network-constrained spatio-temporal clustering analysis of traffic collisions in Jinghan District of Wuhan, China
(By: Yaxin Fan, Xinyan Zhu, Bing She, Wei Guo, Tao Guo)
Aim of Study: To explore the spatio-temporal clustering patterns of traffic collisions by combining a set of network-constrained methods.
Methodology:
- Weighted network kernel density estimation
- Provides an intuitive way to incorporate attribute information
- Network cross K-function
- Shows that there are varying clustering tendencies between traffic collisions and different types of points of interests (POIs)
- Network differential Local Moran’s I and network local indicators of mobility association
- Provides straightforward and quantitative measures of the hotspot changes
Learning Points:
- Taking into consideration of contributory factors that may affect or correlate with the occurrence of traffic collisions (using network cross K-function to quantify spatial interrelationships between two types of point sets, coupled with Monte Carlo simulation method to test the distribution pattern of point events)
- Incorporated semantic information and temporal dimension of traffic collisions for comprehensive understanding of the spatio-temporal clustering patterns
- Considered POIs that might not directly relate to individual collisions but their spatial distribution might correlate with spatial distribution of collisions collectively
Areas for Improvement:
- Perhaps key implementation issues can be highlighted to allow readers to be aware of the potential drawbacks when implementing the mentioned algorithms, and to make necessary changes according to their context.
2. Visualizing Traffic Accident Hotspots Based on Spatial-Temporal Network Kernel Density Estimation
(By: Benjamin Romano, Zhe Jiang)
Aim of Study: Using Spatial-Temporal Network Kernel Density Estimation To Analyse Traffic Accidents in New York
Methodology:
- Spatio-temporal network kernel density estimation (STNKDE)
- In Network Kernel Density Estimation, edges are split into lixels. STNKDE extends the concept of lixel to include a temporal aspect.
Learning Points:
- Use of spatial-temporal methods helps to avoid over-allocation of resources during non-peak hours
Areas for Improvement:
- Explore the use of combining Getis-Ord GI* statistics to determine if hotspots are statistically significant
- Perhaps apply the spatial-temporal aspects to other relevant methods such as Weighted Kernel Density Estimation
3. Analysis of data on the dataset of road accidents in Paris in 2012 and 2013
http://www.remyzum.com/shiny/Paris_Accidentologie/ (By: Remy Zumbiehl)
Aim of Study: Exploratory data analysis of road accident in Paris
Methodology:
- Basic Mapping using Leaflet Package
Learning Points:
- Using different symbol or marker to present the severity of the road accidents
- Adding filters to drill down the dataset
- Types of vehicles
- Temporal Analysis - date range, the day of the week and time frame
Areas for Improvement:
- Add in geospatial analysis such as Kernel Density Estimation and Network Constrained Analysis
APPROACH
After analysing several literature reviews, it is clearly inappropriate to use the standard point pattern analysis used for two-dimensional space. Our analyses need to take into account the geometry of the network and as such, we have to constrain statistical techniques such as K-Function to the network. We will first delve into uncovering which road segments have a higher intensity of road accidents followed by using various network-constrained techniques to investigate if there are statistically significant clusters.
1. Network-Constrained Kernel Density Estimation
Kernel Density Estimation involves estimating the probability density function of a variable, or in other words the density of features in a neighbourhood in geospatial terms. Investigating the average density of points along the network provides a quick insight on which segments of roads has higher intensity of traffic accidents.
There are two main types of intensities when estimating kernel densities in networks:
- Homogeneous intensity function has randomness that is characterised by complete spatial randomness. All points are independent and uniformly distributed in any given set.
- Inhomogeneous intensity function also has independence between disjoint sets but points are unevenly distributed according to their spatially varying intensity functions.
For inhomogeneous intensity function, kernel estimate of intensity has the form:
k(v,u) is the smoothing kernel to smooth out a point on the network, as shown below, in a book by Adrian et al (2016). The smoothing of the point is done with a Gaussian kernel and thickness of line proportional to the kernel value. This can be performed using the density.lpp function in R.
2. Second-Order Network K-Function (Network Constrained Ripley's K-Function)
After gaining insight on the density of traffic accidents, we would take the next step to uncover if the point patterns are dispersed, clustered or randomly distributed. This could be done using Network K-Function.
The Okabe-Yamada network K-Function defines the Network K-Function, by adapting the Ripley’s K Function through replacing the Euclidean distance with the shortest path distance. In this method, given a point v on the network, all locations in the network that can be reached from v by a path of length shorter than a radius r, defined by the user, would be considered. It is defined by the function below, with lambda(L) denoting the total length of the linear network :
While the above Network K-Function constrains points to networks, a second-order stationary point process is required for our analyses. The Okabe-Yamada network K-Function assumes that the network itself is homogeneous, which is not the case as different locations in the network is surrounded by different configurations of line segments. Network K-Functions obtained from different networks are not directly compatible in this case.
The second-order Network K-Function is proposed by Ang et al (2012) , known as the ‘geometrically corrected K-Function’. It is an extension of the Ripley’s K-Function’s benefits of enabling comparison between different point processes with different intensities, observed in different windows, combined with Okabe-Yamada network K-function. The geometrically corrected K-Function is defined by, for all r ≤ R, where u is any location on the network:
The above analysis can be computed in spatstat by the function linearK. It is noted that K-Function assumes homogeneity and Inhomogeneous geometrically K-Function is to be used if the underlying Poisson point process is inhomogeneous.
3. Correlation in Multitype Point Patterns Using Network Cross K-Function
Network K-Function deals with points of the same type while Network Cross-K Function is able to handle two different sets of points. For example, traffic accident points and location of shops could be used and Network Cross-K examines the distribution of accidents in terms of the shortest path from every accident to the nearest shop.
This can be implemented in R through the function linearKcross. The function also has variants for both homogenous and inhomogeneous Poisson point processes.
STORYBOARD
After rounds of refinement, below are the storyboard of our application:
Network Constrained KDE page
Network Constrained KDE page will have 2 maps to show traffic accidents and casualty KDE to allow the user to compare the map. Filters and a kernel distance slider will be at the right-hand side of the application.
Network Constrained K-Function page
Network Constrained K-Function page will have a map at the top and a result box at the bottom which consists of the analysis graph output along with the interpretation function of the graph. Filters and simulation slider will be at the right-hand side of the application.
Network Constrained Cross K-Function and Network Constrained Cross Pair Correlation Function page
Network Constrained Cross K-Function and Network Constrained Cross Pair Correlation Function page will have a map at the top and a result box at the bottom which consists of the analysis graph output along with the interpretation function of the graph. Filters, a variable selection dropdown bar and simulation slider will be at the right-hand side of the application.
TOOLS & TECHNOLOGIES
Tools and technologies
Data Architecture
APPLICATION OVERVIEW
Network Constrained Kernel Density Estimation
Other Network Constrained Analysis
Network Constrained K-Function, Network Constrained Cross K-Function and Network Constrained Cross Pair Correlation Function have the same layout.
Component | Purpose |
---|---|
Leeds Map |
The markers are to indicate the geographical positions of traffic accidents and other variables (Pedestrian Crossing, Motorway Junctions and Schools). The user is able to zoom in and move the map using the map control button to an area that they wish to run the analysis on. |
Kernel Density Estimation Maps |
The user is able to zoom in and move the map using the map control button to an area that they wish to run the analysis on. When the user move or zoom into the top map, the bottom map will be updated automatically as both maps are sync together. This is to ensure that the analysis on both maps are at the same geographical area so that the user is able to have a fair comparison between both maps. |
Time Filters |
The user will be able to filter the data by year, month and hours. |
Environment Filters |
The user will be able to filter the data by:
|
Casualty Filters |
The user will be able to filter the data by:
|
Kernel Distance Slider |
The bandwidth of kernel density plot. The user is able to drag the slider to state the kernel distance in metre which they want to run for the analysis. |
Simulation Slider |
The user is able to drag the slider to state the number of simulations which they want to run for the analysis. |
Variable Selection |
The user is able to choose which variable they want to run the analysis with the traffic accidents. This variable selection is only for Network Constrained Cross K-Function and Network Constrained Cross Pair Correlation Function analysis. |
Network Constrained K-Function |
Output for Network Constrained K-Function Analysis |
Network Constrained Cross K-Function |
Output for Network Constrained Cross K-Function Analysis |
Network Constrained Cross Pair Correlation Function |
Output for Network Constrained Cross Pair Correlation Function Analysis |
Graph Interpretation Function |
The user will choose an option according to the graph output and a general interpretation will be shown on the application. The graph interpretation function is to aid the user to interpret the graph output. This graph interpretation function will be placed in all analysis except Network Constrained KDE. |
INSIGHTS
KEY CHALLENGES
The following are some of the key technical challenges that we may face throughout the course of the project:
Key Challenges | Mitigation Plan |
---|---|
Unfamiliarity with spatial analysis methods related to network constrained |
|
Unfamiliarity with Rshiny |
|
Unfamiliarity with Leeds geographical area |
|
TIMELINE
REFERENCES
- Chua, A. (2018, February 7). Fatal road accidents and fatalities hit all-time low in 2017: Traffic Police. Retrieved from https://www.todayonline.com/singapore/fatal-road-accidents-and-fatalities-hit-all-time-low-2017-traffic-police
- Fan, Y., Zhu, X., She, B., Guo, W. & Guo, T. (2018) Network-constrained spatio-temporal clustering analysis of traffic collisions in Jianghan District of Wuhan, China. PLoS ONE 13(4): e0195093. Retrieved from https://doi.org/10.1371/journal.pone.0195093
- Rui, Y., Yang, Z., Qian T., Khalid, S., Xia N. & Wang J. (2016) Network-constrained and category-based point pattern analysis for Suguo retail stores in Nanjing, China. Retrieved from https://doi.org/10.1080/13658816.2015.1080829
- Turner, A. (2018, November 13). How Big Data is Driving Innovation in the Leeds City Region. Retrieved from https://leeds-list.com/discussion/how-big-data-is-driving-innovation-in-the-leeds-city-region/
- Xie, Z. & Yan, J. (2008). Kernel Density Estimation of traffic accidents in a network space. Computers, Environment and Urban Systems, 32, 396-406. Retrieved from https://doi.org/10.1016/j.compenvurbsys.2008.05.001
- Yamada, I. & Thill, J. (2010). Local Indicators of Network-Constrained Clusters in Spatial Patterns Represented by a Link Attribute. Annals of the Association of American Geographers. 100. 269-285. Retrieved from https://doi.org/10.1080/00045600903550337
- Zumbiehl, R (2018, October 16). Accidentology in Paris. Retrieved from http://www.remyzum.com/shiny/Paris_Accidentologie/
COMMENTS
Feel free to leave us some comments so that we can improve!
No. | Name | Date | Comments |
---|---|---|---|
1. | Insert your name here | Insert date here | Insert comment here |
2. | Insert your name here | Insert date here | Insert comment here |
3. | Insert your name here | Insert date here | Insert comment here |