Conception d’un senseur intégré multimodal pour l’observation des routes

Details

Supervisors

Craeye, Christophe ; Macq, Benoît

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil électromécanicien

Abstract

enThe rise of artificial intelligence, driven by the development of embedded systems, enables the analysis of real and increasingly complex situations. In the field of road observation, the multimodal study of pedestrians and vehicles interactions imposes to choose a detection and classification strategy as well as a multi-target tracking method, based on the synchronized data capture of potentially heterogeneous sensors. In order to record the situations of interaction between pedestrians and vehicles, a device equipped with the appropriate camera and radar is created. Road users are detected and classified by means of tiny-YOLO algorithms, trained with data collected on the site under investigation. Data fusion and target tracking are made possible by a probabilistic data association filter, derived from the Kalman filter. The data collected allow for an efficient training of the tiny-YOLO networks. During the detection and classification stages, an accuracy of 88.1 % and 88.2 % is achieved for the camera and the radar, respectively. The data fusion - which is based on an estimation of the distance by the camera - is a major limitation of the target tracking performance. Despite encouraging results of the data fusion and the target tracking, the implemented algorithm does not seem to be a reliable measure for the study of the interactions between pedestrians and vehicles. The Kalman parameters optimization with an unsupervised learning method is a promising way of improving the obtained results. Nevertheless, the design of a single neural network algorithm, working from the data sensors to target tracking, seems even more promising in the long term.