Estabilización de video en base a la discriminación de puntos de interés utilizando mapas de profundidad.

Abstract

The objective of video stabilization is to optimize the quality of the video, being important to eliminate involuntary movements of the camera, until now the existing methods do not discriminate the points of interest as an initial step for video stabilization. The main objective of this study is to define a video stabilization method that discriminates points of interest using depth maps. It begins by analyzing several methods based on CNN that allow obtaining a robust depth map. Where, the method used by the MiDaS v2.1 trained model is used for the first part of the algorithm that consists of taking each frame and obtaining a depth map. In the second part of the algorithm, a base video stabilization algorithm is defined that receives a sequence of frames as input, takes as input the applied frames a mask defined by the depth map. To achieve video stabilization, we use a validated depth mapping method in combination with a video stabilization method. The algorithm proposed in this work was evaluated in 24 videos in 4 different scenarios. Our algorithm achieves an improvement in the ITF of 5% compared between the metrics of the videos stabilized with the proposed method and the stabilized videos that do not use a depth mask as part of their method.