Segmentação dinâmica de objetos aplicada à odometria visual

Abstract

The presence of dynamic objects in a scene can significantly impair the performance of visual odometry methods. Even with the use of robust methods, it is not always possible to avoid outliers and interferences in the estimation of the camera’s movement. This type of object introduces characteristic points whose movement does not align with the actual movement performed by the camera. To filter these objects, this work presents a neural network architecture that combines RGB images and optical flow to segment regions that exhibit moving objects, even while the camera itself moves. To enable the training of the network, a methodology for quick annotation of object detection datasets is presented to add semantic masks of moving objects to 98,491 images of an urban navigation dataset. The proposed neural network was trained and evaluated with these data and proved adequate for use as a dynamic object filter in visual odometry tasks. To evaluate the proposed model, comparisons of visual odometry algorithms with and without the use of filtering are presented. Based on the results obtained in this work, the identification and filtering of dynamic objects in an image emerges as a fundamental step in the task of visual odometry, being essential for applications involving the presence of dynamic objects.