A three-layer architecture to support disparity map construction in stereo vision systems

Abstract

By adequately locating objects of interest, a computer vision-driven machine can avoid obstacles and collisions, set safe paths to move autonomously, and reconstruct the geometric properties of a scene. In this sense, three-dimensional information is required to perform these tasks properly. As digital images provide information about the scene, they are often used to estimate objects’ spatial position and depths concerning image acquisition machinery. Thus, disparity maps can be obtained, and the differences between similar features from two or more images can be used to encode depth information. Although the disparity map construction has been extensively surveyed, some concepts such as code reuse and compartmentalization need further investigation to optimize application development. Therefore, based on the major stages of disparity calculation, we present a framework to support disparity map construction. This study’s main contributions include a delimitation of elements that integrate stereo vision scope and the modeling of an architecture that integrates the main components of disparity calculation. The proposal is evaluated in different scenarios which demand sparse and dense disparity maps. The standard concepts shared between them are discussed and used to implement a multifaceted application, useful in evaluating stereo vision methods and providing disparity maps as inputs for three-dimensional reconstruction and point cloud elaboration. The code prepared by the authors is publicly available11 https://github.com/pixellab-ufg/disparity-computation-framework.

Keywords

• Software components
• Image processing
• Active vision
• Computer applications
• Robotics