Detecção de pontos-chave em imagens de carros

Abstract

This report details the research conducted on keypoints detection in car bumpers, aimed at establishing an initial stage for a larger system of average speed prediction. To achieve this, an algorithm based on the PyTorch implementation of the algorithm presented in [Sun et al. 2019] was trained, using a dataset collected from aerial images and manually annotated via Labelbox. At the end, a normalized mean error (NME) close to zero was observed. However, due to the inability to apply the solution in scenarios where the vehicle is rotated, an alternative approach is required.