Semantic Neural Network for pedestrian lane detection

Abstract

This work is part of the PG2402ECBTI2022 project approved by the Universidad Nacional Abierta y a Distancia. Its objective is to develop a spatial perception system with artificial intelligence (AI) techniques to improve the orientation and mobility of visually impaired people in Ibagué. The proposed system consists of several modules, and this text focuses on the results of the image processing module oriented to the segmentation of pedestrian lanes through the use of semantic neural networks. The network proposed for this task is based on a Deeplab architecture and was trained on a dataset composed of 5443 images. The results show that the architecture obtained an accuracy level of 0.95 and that its capacity to generalize new samples is in accordance with the task required for the system in general.