Desarrollo de sistema para la detección de baches de la vía Riobamba – El Arenal mediante la aplicación de una cámara estereoscópica y visión por computadora para reducir la probabilidad de caídas bruscas o maniobras de esquiva peligrosas

Abstract

This technical project's objective was to develop a system for detecting potholes on the Riobamba - El Arenal road, applying computer vision and a stereoscopic camera, to reduce the probability of sudden falls or dangerous avoidance maneuvers. With the information obtained from the bibliographical review carried out on pothole detection solutions, the YOLOR-P6 object detection model and the fixed-focus Luxonis OAK-D Pro device was chosen to develop said system. The database was created by capturing color images of 1920x1080 pixel covering 41.27% of the analysis route, on which the pothole and patch classes were labeled, where the latter was included as a filter to avoid erroneous detections, obtaining 630 and 70 images for the training and evaluation of the neural network respectively. This training and evaluation were run on an NVIDIA GeForce RTX 2070 graphics card, and with the resulting weights inferences were run using said graphics card and the Myriad-X mink processing unit, to evaluate its performance. In addition, the records of the accelerometers in the X, Y, and Z axes, generated while driving with and without the pothole detector, were analyzed. In this way, the trained neural network obtained a precision of 49.39% and completeness of 80.09% when evaluating it for both classes, while in its validation for the bump class, a precision of 87.84% was obtained and 67.24% completeness with the Renault Stepway crossover, and 83.73% and 69.87% with the Toyota Corolla Cross SUV. It is concluded that the use of the system allows for reducing the intensity of the fall in a pothole, braking, and lateral dodging maneuvers. It is recommended to strengthen the database with the inclusion of new study routes and climatic conditions.