A decade that got safer, unevenly
Between 2008 and 2017 the police recorded 1,318 collisions in the Municipality of Athens in which a pedestrian was hurt. Most were slight injuries: 1,208 of them. Eighty pedestrians were seriously injured and thirty were killed. In 68 collisions two people on foot were hurt, and in nine, three or four.
The yearly count held at around 180 until 2011, then fell sharply, to 98 in 2013 and 81 by 2017. The economic crisis took cars off the road: car numbers were reported to have fallen by more than 58% in 2011. At the same time, Greece invested in road safety, and in 2021 it received the European Transport Safety Council's award for halving road deaths between 2010 and 2020.
Pedestrian collisions per year, by severity
The more severe collisions fell faster still: 15 killed or seriously injured in 2008, two in 2017. With so few, though, single years jump around, so the maps below use all ten years together.
The streets where it keeps happening
A collision can only happen on a street, so its pattern should be measured along the streets. Each step below changes the map.
Every collision
Each dot is one collision, coloured by how badly the pedestrian was hurt. They follow the big avenues and the dense centre: Patision, Athinas, Peiraios, Alexandras and Syggrou.
Density along the streets
In the thesis I used kernel density, which spreads each collision in a circle across buildings and squares. Here the density runs along the street network instead: every street is cut into 50 m pieces, and each collision counts for the pieces within 250 m of walking distance. The brightest pieces see about seven collisions per kilometre of street per year.
The centre around Omonia
The densest street in Athens is Patision near Omonia and Aiolou: 25 collisions within 150 m in ten years. Athinas, Peiraios and the start of 28is Oktovriou follow close behind. These streets are busy at every hour, lined with shops, and people cross away from the crossings.
Where people were killed or seriously injured
The 109 most serious collisions follow a different map: wide, fast avenues. Alexandras, Vasilissis Sofias and Galatsiou lead, the same pattern the thesis found. Speed decides how badly someone is hurt.
The ten busiest spots
The ten densest stretches, at least 600 m apart, named after the streets of the collisions around them:
Persistent, new and fading hotspots
A hotspot that appears every year needs a different response from one that showed up recently. The thesis used ArcGIS's space-time cube for this. I rebuilt the same logic in Python, on the same 200 m hexagons.
Hotspots, year by year
For each year, every hexagon is tested for being a statistically significant hotspot (Getis-Ord Gi*), and its trend across the decade is checked with a Mann-Kendall test. 20 hexagons are persistent hotspots, hot in at least nine of ten years. Thirteen of them sit around Omonia, matching the thesis.
The north is where it grew
More than half of the new and consecutive hotspots, hot in the last years but not before, are in the northern districts: along Patision towards Galatsiou, and around Kypseli. The thesis flagged these areas as intensifying. They are the places where action now would do the most good.
Clusters and outliers
Local Moran's I asks whether a hexagon looks like its neighbours. 90 high-high cells form the core clusters: a lot of collisions surrounded by a lot of collisions. The few high-low cells are isolated problem spots in otherwise quiet streets.
The centre of gravity barely moves
The mean centre of each year's collisions stays within a few hundred metres, just north-east of Omonia towards Exarcheia, moving slightly north in the later years, as collisions fell faster in the centre than in the northern neighbourhoods.
Ten places I looked at
For the thesis I visited ten of the worst places on Street View: eight persistent and two intensifying hotspots. They have a lot in common: several lanes, traffic lights and crossings, shops, bus stops, and people crossing wherever they can. Each card links to the Street View image used in the thesis. The count is the number of collisions within 150 m in this re-analysis.
Busy places, not bars
The thesis looked at four kinds of places around the collisions: alcohol venues (bars, pubs, clubs), education (schools and colleges), traffic points (crossings, signals, stops) and transport points (bus, tram and metro stops, taxi ranks). The first test is simple: how many of each lie within 100, 250 and 500 m walking distance of a collision?
Points of interest by walking distance from the nearest collision
Measured along the streets, inside the Municipality of Athens
But closeness alone can mislead. The centre has many bars and many people, buses and crossings, so bars would sit near collisions even if they had nothing to do with them. A fairer test counts collisions in each 200 m hexagon, allows for how much street it has, and puts all four kinds of place into one model together. This is a negative binomial regression, the standard model for counts like these.
How much more likely collisions are, for each kind of place
Rate ratio for one standard deviation more of each, with 95% confidence intervals. 1 means no difference.
Once busyness is accounted for, bars and schools no longer explain where collisions happen. Crossings, signals and bus stops do: they mark where people and traffic meet.
That doesn't make crossings dangerous: they are where people cross. It means the risk follows the volume of people meeting traffic, and the fix is in how those meeting points are designed. The thesis then used Geographically Weighted Regression to see where each relationship is strongest. The map below repeats that for traffic points.
Where traffic points matter most
Geographically Weighted Regression gives each hexagon its own rate ratio. Darker hexagons are where a cell with more crossings and signals has many more collisions than its neighbours; hexagons that aren't significant are left blank.
And transport stops
The same for bus, tram and metro stops. The link is significant across much of the city, and most often in the northern districts.
How far help was
The thesis ended with an emergency dashboard: collisions, hospitals, police and fire stations, and Esri's live traffic layer, so a control room could judge how long an ambulance would take. Here the street network does that work. For every collision I measured the drive along the streets to the nearest public general hospital, police station and fire station, at an average of 30 km/h and with no traffic.
Athens has a particular complication: public hospitals take turns on emergency duty (εφημερίες), so the nearest hospital that will take a patient changes from day to day. The dashboard lets you tick which hospitals are on duty and recalculates every time on the map.
Minutes to the nearest hospital
Streets coloured by the drive to the nearest public general hospital, assuming all are open. The eastern districts, near Evangelismos, Ippokrateio, Laiko and Elpis, are within a few minutes. The far west, around Kolonos, Sepolia, Iera Odos and Athinon Avenue, is furthest away: a median of about 8 minutes, against under 1 minute in the east.
When only one hospital is on duty
If Evangelismos were the only hospital on duty, collisions in the far west of the municipality would be over 10 minutes away even without traffic. That's why the rota matters, and why live traffic would make it worse.
Fire and rescue
OpenStreetMap shows only two fire stations in or near the municipality, so these times are an upper limit. If the data is incomplete, the real times are shorter.
Try it yourself. Click anywhere in Athens to see the nearest on-duty hospitals, police and fire stations, and the collision history around that spot.
Open the dashboard →What the evidence points to
The thesis closed with three observations that this re-analysis supports:
The meeting points need redesigning. The persistent hotspots are wide streets with many lanes, where crossings sit far apart and people cross between them. Narrower crossings, refuge islands, longer green time for pedestrians and lower speeds on the avenues address where the risk actually is.
The north deserves attention now. While the centre improved, new and consecutive hotspots appeared around Patision, Galatsiou and Kypseli.
Culture matters too. Drivers often don't give way, and jaywalking is common. Education and enforcement work alongside design, not instead of it.
What the data can't say: it has only the year of each collision, not the day or hour, nor the age of the pedestrian or the type of vehicle. With those, the same methods could tell night-time collisions near bars from school-run collisions, as my London project does with British police data.
From ArcGIS in 2021 to Python in 2026
Thesis, 2021
- ArcGIS Pro and ArcMap (student licence)
- Planar kernel density
- Space-time cube (emerging hotspots)
- Anselin Local Moran's I
- Network Analyst walking service areas
- Bivariate maps and GWR
- ArcGIS Online dashboard with live traffic
This version, 2026
- Python: GeoPandas, NetworkX, PySAL, statsmodels, mgwr
- Network kernel density along the streets
- Emerging hotspots rebuilt (Gi* per year + Mann-Kendall)
- Local Moran's I, 999 permutations
- Walking distances along the network
- Negative binomial with street length as exposure, then GWR
- Open dashboard with an on-duty hospital rota; live traffic optional
Everything is reproducible from the thesis data with five scripts, and the code is on GitHub. Distances are best-case: the thesis street network has no one-way information, so routes may use streets in either direction. Hospitals, police and fire stations come from OpenStreetMap; public hospitals that OpenStreetMap stores only as building outlines are placed at the bus stop named after them, by their entrance.