Noemi Onofa’s scientific contributions

This article presents an algorithm for the detection of pedestrians in urban driving environments during the day. The main contribution is in the design of a new classifier to discriminate between the person and the background, under partial occlusion. To construct the classifier, the HOG (Histogram of Oriented Gradients) descriptor was used together with the SVM (Support Vector Machine) and IL (Logic Inference) algorithms.The input image has been divided into twelve regions, and for each of them the feature vector has been extracted and a classifier based on SVM has been built. With this design it is possible to capture the specific detail of each part of the human body, such as head, legs, arms and body. Subsequently, they have been joined in a final classifier using IL, in order to obtain an efficient algorithm to discriminate between partially occluded pedestrians and the background, in urban environments during the day. The experiments related to the classifier were developed onseveral public databases, in various degrees of partial occlusion; and the experiments linked to the detection were generated on the visual information obtained by the experimental platform ViiA, to validate the proposal under real driving conditions.

Este artículo presenta el desarrollo de un algoritmo para la generación de regiones de interés con alto potencial de contener peatones sobre imágenes monoculares. Para la generación de estas regiones se ha construido un algoritmo de generación de hiperplanos de búsqueda en función de la carretera junto con la generación de ventanas aleatorias, sobre ésta zona con una variación de la técnica desplazamiento de ventana piramidal; luego se realiza el pre-procesamiento usando los filtros de gradientes vertical y horizontal. Para comprobar que la región es un posible peatón, se parte de dos hipótesis respecto a la figura humana, la componente vertical es mayor que la horizontal y la fuerte simetría vertical. Mediante este proceso se obtiene un conjunto reducido y óptimo de regiones en el intervalo de entre 2 y 25 metros al frente de la cámara. Los resultados experimentales, desarrollados sobre las bases de datos presentes en el estado del arte, muestran que se tiene una tasa del 91% de ventanas válidas con respecto al total de ventanas verdaderas, a 25.38 fotogramas por segundo.