Detecção e estimação de distância de marcos visuais por um veículo autônomo a partir de segmentação de imagens com aprendizado profundo e processos Gaussianos

• Main Author: Bernardes, Danilo Serenini
• Publication Date: 2024
• Format: Master thesis
• Language: por
• Source: Repositório Institucional da UFLA
• Download full: https://repositorio.ufla.br/handle/1/59313
• topic: Veículos autônomos; Visão computacional; Deep learning; Detecção de imagens; Segmentação de imagens; Regressão por processos Gaussianos; Estimação de distância; Aprendizado de máquina; Autonomous vehicles; Computer vision; Image detection; Image segmentation; Gaussian process regression; Distance estimation; Machine learning; Engenharia

In the constantly evolving scenario of technologies for autonomous vehicles implementation, precision in vehicle localization emerges as a significant challenge. The objective of this work is to propose an algorithm that applies distance prediction techniques to estimate the distance between a vehicle equipped with a camera and landmarks in the environment it will be exposed to. For this purpose, computer vision techniques were employed for landmark detection, followed by object segmentation to enhance the algorithm's perception of the environment. For the development of the prediction algorithm, the Python language was chosen, and a real database with approximately 8000 samples collected in the field at the University of Waterloo, through an instrumented autonomous vehicle, was considered. The use of YOLO-v8 network with Object Detection and Segmentation models, DeTr (Detection Transformers) network, and SAM (Segment Anything Model) network were evaluated to provide the input data that were related to distance estimation from deep learning techniques with a GPR (Gaussian Process Regression) model. At the end of the project, the superiority of YOLO-v8 network in the segmentation model when applied to object detection task was observed, with an average Recall of 0.76 and a maP@0.5 of 0.891, highlighting the benefit of using segmentation masks also for object detection. The analysis showed that the combination of YOLO-v8 Segmentation and SAM networks enhances the environment perception with a DICE coefficient of 71.039% and significantly reduces the error in distance prediction, achieving a MAE (Mean Absolute Error) of 0.65 meters. However, this combination resulted in an increase in processing time, standing out as a challenge for real-time application from the perspective of the hardware used. From the results, it is possible to note that the incorporation of segmentation characteristics to the input data substantially improves the performance of the GPR model in distance prediction, highlighting the potential of computer vision techniques in improving the localization and decision-making in autonomous vehicles.