A Fast Feature Tracking Algorithm for Visual Odometry and Mapping Based on RGB-D Sensors

Abstract

The recent introduction of low cost sensors such as the Kinect allows the design of real-time applications (i.e. for Robotics) that exploit novel capabilities. One such application is Visual Odometry, a fundamental module of any robotic platform that uses the synchronized color/depth streams captured by these devices to build a map representation of the environment at the same that the robot is localized within the map. Aiming to minimize error accumulation inherent to the process of robot localization, we design a visual feature tracker that works as the front-end of a Visual Odometry system for RGB-D sensors. Feature points are added to the tracker selectively based on pre-specified criteria such as the number of currently active points and their spatial distribution throughout the image. Our proposal is a tracking strategy that allows real-time camera pose computation (average of 24.847 ms per frame) despite the fact that no specialized hardware (such as modern GPUs) is employed. Experiments carried out on publicly available benchmark and datasets demonstrate the usefulness of the method, which achieved RMSE rates superior to the state-of-the-art RGB-D SLAM algorithm.

Efficient 3D Objects Recognition Using Multifoveated Point Clouds

“…According to the literature [7,8,30,33], we used this approach to evaluate the performance of the descriptors and foveated models.…”

“…Thus, it is an instance classification problem, i.e., a feature association problem. According to Da Silva [8], the classification problem can have the following classifications for the associated features: true positive association that was established correctly, false positive association established incorrectly, true negative association not established correctly, and false negative association not established correctly. However, in this proposal, we do not have exclusive interest in the classification of the key-points by the descriptors, since these points were fully discussed in previous works [7,30,33] in relation to the 3D descriptors with focus in SHOT.…”

“…Since this is a key-point correspondences problem (local features), we complement our analysis using an approach assembling comparative graphs [8] that address and emphasize the precision and sensitivity of the chosen methods. This is done based on ground-truth (estimates considered true, i.e., a model to be followed) provided by [51] and Tombari [52].…”

“…This simplifies the layout of the approaches in the exposed graphs, seen later. As it is not a question of evaluating only the descriptor capacity, but also the ability of the models to perform the task of recognizing a model in a point cloud, we do not use the $1-precision\times recall$ curve used by Silva and Gonçalves [8], Salti et al [30], and Tombari et al [33]. …”

“…To achieve high level of autonomy, it is necessary for a robot to react without human intervention to stimuli provided by the environment [8,9,10]. Robotic vision applications use several types of sensors to capture stimuli data.…”