Urban intersection safety risk index - Machine learning methods for real-time classification

Aug 5, 2026

Urban intersection safety is a central concern for planners, and modern sensing such as LiDAR enables more data-driven risk assessment. This study develops a hybrid framework that uses Post-Encroachment Time (PET) combined with unsupervised machine learning, specifically DBSCAN clustering, to detect traffic anomalies. A generalized Pareto distribution estimates a risk index, followed by categorical safety classification via ONN, OSVM, ELR, and GNB. Real-time evaluation supports traffic management and accident prevention in Trois-Rivières, Quebec, Canada.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages5

Reading Time13 min

AccessFree to download

Introduction

Method overview and contributions

Methodology

PET calculation

Data description

# Urban intersection safety risk index:Machinelearning methods for real-time classification Thierno Fall Daniel Massicotte Departement of Electrical and Computer EngineeringDepartement of Electrical and Computer EngineeringUniversité du Québec à Trois-RivieresUniversité du Québecà Trois-RivieresTrois-Rivieres,CanadaTrois-Rivières,CanadaMame.Thierno.Mbacke.Fall@uqtr.caDaniel.Massicotte@uqtr.ca Messaoud Ahmed Ouameur Jean-Sébastien Dessureault Departement of Mathematics and Computer scienceDepartement of Electrical and Computer EngineeringUniversité du Québec à Trois-RivièresUniversité du Québecà Trois-RivièresTrois-Rivieres,CanadaTrois-Rivières,Canadajean-sebastien.dessureault@uqtr.camessaoud.ahmed.ouameur@uqtr.ca PET value less than zero would indicate a crash occurrence[3]. Abstract—The safety of urban intersections is a criticalconcern for city planners.Technological advancements,suchas LiDAR sensors,enable better risk assessment for roadusers.This study proposes a hybrid model that combines Post-Encroachment Time(PET)data with unsupervised machinelearning techniques,specifically DBSCAN clustering,to detecttraffic anomalies.A generalized Pareto distribution(GPD)sthen applied to estimate a risk index.Finally,categorical safetyrisk classification is performed using an optimizable neuralnetwork(ONN),support vector machine(OSVM),efficientlogistic regression(ELR),and Gaussian Naive Bayes(GNB).Theimpact of these methods is evaluated in real-time for urban trafficmanagement in Trois-Rivieres,Quebec,Canada.This work aimsto assist decision-makers in urban traffic planning and accidentprevention. Due to the rarity of dangerous events,such as accidentsand near-misses,and the lack of timeliness,having a reliablemodel that gives a safety index of road traffic is challenging.Surrogate traffic measures can be used to determine howdangerous an intersection or road section is. [4]proposes a conflict-based traffic safety assessmentmethod by associating conflict frequency and severity withshort-term traffic characteristics.[5]conducts a systematicreview of conflict-based safety measures with a specific focuson the context of their applications.[6]uses conflict indica-tors,PET and Time to Collision(TTC)to identify pedestrianconflicts and predict pedestrian conflicts one cycle ahead,which can be 2-3 min,whereas [7]employed surrogate safetyindicators to measure the safety level of pedestrian conflictwith other road users to evaluate conflict risk. Index Terms—Surrogate measures,Post-Encroachment Time(PET),Machine learning,DBSCAN,Pareto distribution,Classi-fication. ## I.INTRODUCTION Despite the significant increase in interest in surrogatemeasures,many approaches lack real-time applicability.Thisstudy introduces a machine learning-based risk index capableof real-time classification of intersection safety levels.Ourcontribution uses PET as a surrogate measure and focuses on Intelligent transport systems develop and integrate methodsto overcome the high demand for economic concerns,thereliability and quality of infrastructures,and most importantly,the safety of road users.Each year,1.35 million people arekilled on the roads of the world,and another 20 to 50 millionare seriously injured [1]. 1)machine learning techniques to evaluate the behavior ofindividual road intersections,2)calculating the safety indexfor an hourly divided block of conflicts of all the data framesfrom ten intersections,and 3)using the generated safety risklevel and conflict features(PET and speed)to train and testclassification methods. In road traffic,many users interact with each other,and theneed for reliable measurement systems (on a week,day,andeven hour basis)of interactions between users is growing.These interactions will likely generate conflicts(betweenvehicles,pedestrians,and motorcycles)as they cross pathsin all directions.Thus,conflicts occur when traffic streamsmoving in different directions interfere.The number of possi-ble conflict p

FAQs

Why is a real-time urban intersection safety risk index challenging to build?

Dangerous events are rare and near-misses occur infrequently, while timely models are difficult to maintain. The paper motivates surrogate measures to quantify risk despite limited dangerous observations.

How does the proposed method use PET for risk assessment?

PET is defined as the time gap between when a first road user exits a conflict point and when a second user enters the same conflict spot. Smaller PET values correspond to higher collision risk.

Which models are used for anomaly detection and safety classification?

DBSCAN clustering is applied for traffic anomaly detection using PET-based data. Safety risk is then classified categorically using ONN, OSVM, ELR, and Gaussian Naive Bayes after estimating the risk index with a generalized Pareto distribution.

Aug 5, 2026

Upload by Clementine