Year: 2024
Type: conference-paper
Abstract: The need for redundant odometry in robotic systems benefits from the recent possibilities of fast and efficient real-time image processing using classifications and regression models. The present work proposes an approach using Convolutional Neural Networks (CNN) to detect the distance between the point of capture of an image and a water surface below the capture. Our CNN was trained initially with synthetic images generated by Blender, and evaluated by real images at heights between 50 and 100 cm. Preliminary tests demonstrated an ability of the system to recognize the changing of height in a continuous shot, varying only the vertical position, even with some noise from elements not present in the training test database, such as dirt from leaves and stones at the bottom of the water.
Authors: João Vitor da Silva Neto, Tiago Pereira do Nascimento, Leonardo Vidal Batista
Institution: Universidade Federal da Paraíba
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4408888869