💻 computer science

Multi-Scale Unsupervised Change Point Detection in Long-Term Urban Traffic Time Series: A Kernel-Based Framework Applied to Rotterdam

This paper presents a generic unsupervised, kernel-based framework that combines macroscopic and microscopic traffic indicators to automatically detect and spatially localize both major network-wide disruptions and subtle, localized behavioral shifts in long-term urban traffic time series, as demonstrated by a seven-year analysis of Rotterdam's data.

Original authors: Géraldine Viel, Irene Martínez, Ludovic Leclercq

Published 2026-09-28

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Géraldine Viel, Irene Martínez, Ludovic Leclercq

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Cities are living systems, constantly breathing in and out with the rhythm of their inhabitants. For decades, traffic engineers have relied on loops of wire buried beneath the asphalt to count cars and measure speed, creating a vast, continuous record of how people move. Yet, much of this data sits dormant, treated as a static archive rather than a dynamic story. Planners often look at traffic in two ways: either they zoom in so closely on a single street that they miss the bigger picture, or they zoom out so far to look at the whole city that they blur the details of local changes. The result is a gap in understanding. We know when a major event, like a pandemic, stops traffic, but we often miss the quieter, more subtle shifts that happen when a bridge is repaired, a speed limit changes, or a neighborhood slowly adapts to new habits. Without a way to see these moments clearly, it is difficult to know if a city is truly changing or just fluctuating with the weather.

A team of researchers has developed a new way to listen to the city's heartbeat, one that can hear both the thunderous disruptions and the faint whispers of change. They created a method to scan years of traffic data and automatically pinpoint the exact dates when the behavior of the road network shifted in a lasting way. Instead of guessing when to look, their system scans the entire timeline, from the broad patterns of the whole city down to the specific conditions on individual road segments. It does this by comparing how traffic flows today against how it flowed yesterday, looking for moments where the rhythm broke and a new pattern took hold. The researchers applied this method to seven years of data from Rotterdam, a major port city in the Netherlands, to see what stories the road network had been telling all along.

The approach works like a careful, layered investigation. First, the researchers built a set of daily scores that describe the health of the entire road network. These scores measure things like how fast cars move when the roads are empty, how much traffic the network can handle before it clogs, and how long the rush hours last. They also looked at the specific conditions on individual roads, but with a clever twist: instead of just looking at the raw speed on one street, they tracked how that street's performance ranked compared to every other street in the city. This allowed them to see if a specific road was becoming relatively slower or faster than the rest of the network, a sign of a local problem or improvement that might be hidden in a city-wide average.

Once these daily scores were ready, the researchers used a computer algorithm to find the moments when the scores changed permanently. The algorithm was designed to ignore short-term noise, like a temporary holiday or a school break, and focus only on shifts that lasted for at least two months. It worked in a top-down fashion, first looking for changes that affected the whole city at once. If it found a major shift, it marked that date and then looked for smaller, more specific changes that happened between those major dates. This ensured that the system didn't get confused by the same big event showing up multiple times. Finally, it examined the individual roads to see if any specific segments had changed their behavior relative to the rest of the network.

When applied to Rotterdam, the system produced a clear timeline of the city's evolution between 2016 and 2023. The most obvious change appeared in March 2020, when the network suddenly collapsed as the pandemic lockdown began. The algorithm correctly identified this as a massive, city-wide disruption that affected every metric it measured. But the system also found changes that a simple glance at the data would have missed. For instance, it detected a shift in the summer of 2019 that coincided with the full reopening of the Maastunnel, a major river crossing that had been under construction for two years. The data showed that once the tunnel reopened, the flow of traffic across the river reorganized, changing the capacity of the entire network.

Another significant finding occurred in April 2021, when the city's main central boulevard, the Coolsingel, reopened after a three-year reconstruction that had reduced its lanes. The algorithm pinpointed the exact date when the free-flow speed on the network changed, confirming that the permanent reduction in road capacity had a lasting impact. The system also caught a subtle but rapid shift in early 2022, just after the government lifted work-from-home mandates. While the overall city traffic didn't show a massive spike, the data revealed that the flow of commuters reorganized almost overnight, with different corridors absorbing the returning workers in uneven ways. This was a change so specific to the local flow of people that it would have been invisible to a study looking only at city-wide averages.

The power of this method lies in its ability to connect the dots between the big picture and the local details. By tracking how individual roads changed their rank compared to their neighbors, the researchers could see exactly where the impact of a change was felt. For example, they found that a road near the Maastunnel, which had been heavily congested during the construction, became significantly more resilient to traffic jams once the tunnel reopened. Conversely, they also found that some major construction projects, like a junction renovation in 2023, did not leave a detectable mark on the network's overall behavior, suggesting that the changes were either too gradual or too localized to shift the city's main traffic patterns.

This work suggests that cities are constantly rewriting their own stories, often in ways that are too subtle for traditional analysis to catch. The researchers did not claim to have solved the mystery of why every change happened, nor did they prove that their method is perfect for every city. Instead, they provided a tool that can automatically generate a dated and scored record of when and where the traffic network changed. This record gives planners a systematic way to audit their data, evaluate the long-term effects of past projects, and identify stable periods that are suitable for further study. By turning years of raw numbers into a clear narrative of structural shifts, the method offers a new way to understand the complex, evolving life of an urban road network.

Technical Summary: Multi-Scale Unsupervised Change Point Detection in Long-Term Urban Traffic Time Series

Problem Statement
Urban road networks generate extensive long-term traffic datasets, yet their full spatio-temporal potential remains under-exploited. Existing analytical approaches suffer from two primary limitations:

Over-aggregation: Studies focusing on network-wide behavior often aggregate data over long periods or entire networks, obscuring spatial granularity and missing localized, unanticipated changes.
Event-driven bias: Many retrospective studies rely on pre-identified events (e.g., specific infrastructure projects) to perform before-after comparisons, failing to detect unanticipated structural shifts or subtle behavioral reorganizations.
Lack of systematic frameworks: While major events like the COVID-19 pandemic have been studied extensively, no single methodology exists to systematically detect and date all structural changes across a network without prior assumptions, covering both global disruptions and local reorganizations.

Methodology
The authors propose a generic, unsupervised, two-stage framework applied to loop detector and probe vehicle data. The workflow is hierarchical and top-down, designed to filter noise and seasonality while isolating genuine structural shifts.

Stage 1: Performance Indicator Generation
The framework computes daily time series at two complementary scales:
- Macroscopic Scale (Network-wide): Indicators aggregate measurements to characterize global network state. These include:
  - MFD-inspired parameters: Free-flow speed ( $V_f$ ), maximum flow ( $Q_{max}$ ), critical density ( $K_c$ ), and curvature indicators ( $\Delta_s$ ) derived from piecewise linear models of flow-speed relationships.
  - Demand and congestion metrics: Normalized Total Travel Length ( $TTL_{norm}$ ), peak hour duration ( $D^{AM}_{peak}, D^{PM}_{peak}$ ), peak intensity ratios ( $\eta$ ), and speed spatial heterogeneity ( $\sigma_v$ ).
- Microscopic Scale (Local): Indicators are computed for individual road segments. Crucially, rather than tracking raw values, the method tracks the rank of each detector within the network distribution (e.g., quartile membership). This distinguishes local changes in a segment's relative importance from network-wide demand shifts. Indicators include extremum traffic states ( $v_{10}^{peak}, q_{90}$ ), free-flow speed distribution percentiles, and loop-specific quartile memberships ( $Q_{l,t}$ ).
- Preprocessing: A Hampel filter is applied to remove non-physical outliers (sensor errors) while preserving structural changes.
Stage 2: Hierarchical Change-Point Detection (CPD)
A kernel-based CPD algorithm (KernelCPD with an RBF kernel) is applied using a four-step sequential workflow. Change points detected in earlier steps are imposed as fixed boundaries in subsequent steps to prevent redundant detection of the same event.
1. Joint Macroscopic: Simultaneous analysis of all macroscopic indicators to detect major, network-wide disruptions.
2. Individual Macroscopic: Univariate analysis of each macroscopic indicator to detect dimension-specific shifts masked during joint analysis.
3. Joint Microscopic: Simultaneous analysis of microscopic distribution indicators (excluding the high-dimensional quartile series) to detect widespread spatial reorganizations.
4. Individual Microscopic: Univariate analysis of each loop's quartile membership series to identify precise, localized changes in specific road segments.
Impact Scoring: Each detected change point is assigned an objective impact score defined as the relative reduction in the segmentation cost function, allowing for cross-scale comparison of event intensity.

Key Contributions

A Generic Unsupervised Framework: The paper introduces a methodology that does not require prior knowledge of data distributions or pre-identified events, capable of operating on any network with available sensor data.
Multi-Scale Integration: By combining macroscopic aggregation with microscopic rank-based analysis, the framework captures both global structural breaks and subtle, spatially localized reorganizations that would be invisible to network-wide averaging.
Hierarchical Detection Strategy: The top-down approach (Joint $\to$ Individual, Macro $\to$ Micro) ensures that dominant events are identified first, allowing the algorithm to focus subsequent sensitivity on finer, subtler transitions within stable periods.
Objective Impact Quantification: The use of a normalized impact score provides a systematic way to rank detected events by magnitude.

Results: Rotterdam Case Study (2016–2023)
The framework was applied to seven years of loop detector data from Rotterdam.

Macroscopic Findings: The analysis successfully identified major network-wide disruptions, most notably the onset of the COVID-19 lockdown on March 13, 2020 (impact: 11.08%). It also detected dimension-specific shifts, such as changes in free-flow speed ( $V_f$ ) coinciding with the reopening of the Coolsingel (April 2021) and changes in flow capacity ( $Q_{max}$ ) linked to the Maastunnel reopening (June–July 2019).
Microscopic Findings: The framework identified changes invisible to global aggregation. For example, a cluster of breakpoints in late January/February 2022 was detected only at the individual microscopic level, attributed to the abrupt return of office workers following the lifting of Dutch work-from-home mandates. This event showed spatial heterogeneity (uneven demand redistribution) that did not trigger a network-wide signal.
Localization: The loop-level quartile analysis allowed for the precise mapping of specific segments affected by events. For instance, the Westzeedijk loop showed a persistent shift to higher speed quartiles following the Maastunnel works, indicating increased resilience relative to the network.
Negative Results: The framework also identified the absence of signals for certain events (e.g., the Roseknoop junction renovation in 2023), helping to characterize the sensitivity threshold of the method.

Significance and Claims
The paper claims that this framework offers urban practitioners a systematic tool to:

Audit Traffic Data: Automatically generate a dated and impact-scored record of network behavioral shifts.
Evaluate Interventions: Assess the effect of past interventions by identifying structural breaks that correlate with specific dates, serving as a dependent variable for future causal inference studies.
Identify Stable Periods: Delineate periods of stable network behavior suitable for further modeling.

The authors emphasize that while the framework successfully detects and dates changes, it does not establish causal links; formal causal inference requires a separate step comparing the affected network against a counterfactual. Furthermore, the method quantifies the magnitude of change but does not inherently characterize the nature (e.g., increase vs. decrease) of the shift without further interpretation. The study concludes that the proposed method addresses a longstanding limitation in urban traffic analysis by fully exploiting the temporal depth and spatial granularity of existing loop detector archives.

More like this