Phd in INFRASTRUCTURES AND TRANSPORT

Thesis title: Traffic conflict approach for road safety analysis : New insights and development of conflict-based safety estimation models for mixed traffic intersection in LMICs, leveraging artificial intelligence-based video analytics

Road traffic fatalities and injuries remain a major global concern, disproportionately affecting low- and middle-income countries (LMICs), where the majority of crashes occur despite limited data and institutional capacity to manage road safety effectively. In these settings, reliance on crash-based approaches constrains proactive safety management due to the rarity, underreporting, and delayed availability of crash data. Traffic conflict-based safety analysis offers a proactive and data-efficient alternative by using near-miss interactions as surrogates for crash risk. However, its application in LMICs remains limited, largely due to methodological challenges, lack of validation under heterogeneous, non-lane-based traffic conditions, and the absence of context-specific models and thresholds suited to mixed traffic environments dominated by vulnerable road users (VRUs). This thesis addresses these critical gaps through two main objectives: (1) to identify, synthesize, and improve insights and knowledge on the conceptual, methodological, and practical foundations of conflict-based safety analysis, particularly in LMICs; and (2) to develop, calibrate, and validate conflict-based safety estimation models for mixed-traffic intersections in Africa capable of predicting hourly conflict frequency and annual crash frequency by severity. To achieve these objectives, a multi-method approach combining systematic literature reviews, empirical data collection, and advanced statistical modelling was adopted. The first objective was addressed through exploratory, scoping, and systematic reviews that synthesized over 300 studies, covering conceptual foundations of traffic conflicts, data collection techniques, video-based processing methods, surrogate safety indicators (SSIs), threshold-setting approaches, modelling strategies, and validation techniques. The second objective involved large-scale data collection from 22 intersections in Cameroon and Italy, selected for their geometric diversity, traffic mix, and crash history. Using advanced artificial intelligence (AI)-based video analytics, specifically YOLO (CNN) for detection, ByteTrack for tracking, and Tsai for calibration, road user trajectories were extracted and processed through calibration and homography transformation to compute SSIs such as Time-to-Collision (TTC), Post-Encroachment Time (PET), and Delta-V. The results validated the suitability of TTC and PET as reliable SSIs for mixed-traffic intersections, with an empirically derived critical TTC threshold of 2.0 s for distinguishing safety-critical events. Negative Binomial (NB) models revealed that traffic exposure and modal heterogeneity were the primary determinants of conflict frequency, while geometric attributes exerted secondary, site-specific influences. A novel heterogeneity index (H) was developed and shown to significantly affect conflict occurrence, with non-linear relationships varying by conflict type. For crash estimation, Extreme Value Theory (EVT) models, both Block Maxima (BM) and Peak Over Threshold (POT), were applied to proximity (TTC) and severity (ΔV) indicators, incorporating non-stationary covariates such as speed and heterogeneity. The bivariate POT-GPD model achieved superior performance, accurately reproducing observed crash frequencies and severities, thereby validating the predictive capacity of conflict-based models for LMIC contexts. This research makes several key contributions: it provides a validated framework for the use of conflict-based safety analysis in LMICs; establishes context-specific SSI thresholds; develops the first conflict-based crash prediction models in Africa accounting for non-stationarity; and demonstrates the value of AI-driven video analytics for proactive safety assessment. Beyond road safety applications, the developed models can support infrastructure evaluation, policy prioritization, and socio-economic appraisal of safety interventions. Overall, this thesis advances the scientific foundation and practical feasibility of conflict-based road safety analysis, providing an evidence-based framework to strengthen proactive safety management and policy decision-making in LMICs. Keywords: Traffic conflicts, surrogate safety indicators, mixed traffic, intersections, video analytics, extreme value theory, crash prediction, heterogeneity, vulnerable road users, low- and middle-income countries.