Thesis Evaluación comparativa de la familia de modelos YOLO11 para analítica de tráfico urbano
No Thumbnail Available
Date
2026-10-06
Authors
Journal Title
Journal ISSN
Volume Title
Program
Ingeniería Civil Industrial
Departament
Campus
Campus Casa Central Valparaíso
Abstract
La gestión del tránsito urbano depende de saber qué circula por cada intersección y en qué momento, y las cámaras ya instaladas en las ciudades pueden entregar ese dato si se les suma un detector de objetos capaz de operar en tiempo real. Elegir cuál detector desplegar, sin embargo, suele resolverse mirando solo la precisión sobre un conjunto de prueba, sin preguntarse cuánto vale esa precisión adicional una vez puesta en la calle. Esta memoria evalúa los cinco tamaños de la familia YOLO11 (n, s, m, l y x) sobre BDD100K y traduce su desempeño técnico a valor social, con el fin de determinar cuál ofrece el mejor equilibrio entre calidad y costo. Los cinco modelos se ajustaron bajo un protocolo idéntico sobre un subconjunto de 15.000 imágenes de entrenamiento y 3.000 de validación, cada una etiquetada por momento del día, clima y densidad vehicular, lo que permitió medir el desempeño dentro de cada condición en lugar de un promedio que las mezcla. La robustez se estimó por ponderación de probabilidad inversa y por comparaciones con las variables de confusión fijadas, y el costo se midió con telemetría de inferencia sobre GPU. Ese desempeño se llevó a una valorización social sobre las cerca de 3.250 intersecciones semaforizadas de la Región Metropolitana, con precios sociales y una tasa de descuento de 5,5%, valorizando el ahorro de tiempo por menor congestión, los siniestros evitados y las emisiones de CO₂ evitadas, más un análisis de riesgo por simulación de Montecarlo de 100.000 iteraciones sobre 18 parámetros. El mAP50-95 crece de 0,244 en YOLO11n a 0,338 en YOLO11x y se estanca a partir de YOLO11m, cuyo intervalo de confianza al 95% se solapa con el de los dos modelos mayores. La noche resulta ser la condición crítica para los cinco tamaños, con caídas de entre (...).
Urban traffic management depends on knowing what circulates through each intersection and at what time, and the cameras already installed across cities can provide that data once an object detector capable of running in real time is added to them. Choosing which detector to deploy, however, is usually settled by looking only at accuracy on a test set, without asking how much that additional accuracy is worth once the system is on the street. This thesis evaluates the five model sizes of the YOLO11 family (n, s, m, l and x) on BDD100K and translates their technical performance into social value, in order to determine which one offers the best balance between quality and cost. The five models were fine-tuned under an identical protocol on a subset of 15,000 training images and 3,000 validation images, each labelled by time of day, weather and vehicle density, which made it possible to measure performance within each condition rather than as an average that blends them. Robustness was estimated through inverse probability weighting and through comparisons with the confounding variables held fixed, while cost was measured with GPU inference telemetry. That performance was then carried into a social appraisal covering the roughly 3,250 signalized intersections of the Santiago Metropolitan Region, using social prices and a 5.5% discount rate, valuing the travel time saved through reduced congestion, the road crashes avoided and the CO₂ emissions avoided, together with a risk analysis based on a Monte Carlo simulation of 100,000 iterations over 18 parameters. The mAP50-95 rises from 0.244 for YOLO11n to 0.338 for YOLO11x and levels off from YOLO11m onwards, whose 95% confidence interval overlaps with those of the two larger models. Night proves to be the critical condition for all five sizes, with drops of between (...).
Urban traffic management depends on knowing what circulates through each intersection and at what time, and the cameras already installed across cities can provide that data once an object detector capable of running in real time is added to them. Choosing which detector to deploy, however, is usually settled by looking only at accuracy on a test set, without asking how much that additional accuracy is worth once the system is on the street. This thesis evaluates the five model sizes of the YOLO11 family (n, s, m, l and x) on BDD100K and translates their technical performance into social value, in order to determine which one offers the best balance between quality and cost. The five models were fine-tuned under an identical protocol on a subset of 15,000 training images and 3,000 validation images, each labelled by time of day, weather and vehicle density, which made it possible to measure performance within each condition rather than as an average that blends them. Robustness was estimated through inverse probability weighting and through comparisons with the confounding variables held fixed, while cost was measured with GPU inference telemetry. That performance was then carried into a social appraisal covering the roughly 3,250 signalized intersections of the Santiago Metropolitan Region, using social prices and a 5.5% discount rate, valuing the travel time saved through reduced congestion, the road crashes avoided and the CO₂ emissions avoided, together with a risk analysis based on a Monte Carlo simulation of 100,000 iterations over 18 parameters. The mAP50-95 rises from 0.244 for YOLO11n to 0.338 for YOLO11x and levels off from YOLO11m onwards, whose 95% confidence interval overlaps with those of the two larger models. Night proves to be the critical condition for all five sizes, with drops of between (...).
Description
Keywords
Detección de objetos, Visión computacional, Analítica de tráico urbano, Semáforos adaptativos, Evaluación social de proyectos, Simulación de Montecarlo, Object detection, Computer vision, Urban traffic analytics, Adaptive traffic signals, Social costo-benefit analysis., Monte Carlo simulation
