We explore the problem of real-time stereo matching on high-res imagery. Many state-of-the-art (SOTA) methods struggle to process high-res imagery because of memory constraints or speed limitations. To address this issue, we propose an end-to-end framework that searches for correspondences incrementally over a coarse-to-fine hierarchy. Because high-res stereo datasets are relatively rare, we introduce a dataset with high-res stereo pairs for both training and evaluation. Our approach achieved SOTA performance on Middlebury-v3 and KITTI-15 while running significantly faster than its competitors. The hierarchical design also naturally allows for anytime on-demand reports of disparity by capping intermediate coarse results, allowing us to accurately predict disparity for near-range structures with low latency (30ms). We demonstrate that the performance-vs-speed tradeoff afforded by on-demand hierarchies may address sensing needs for time-critical applications such as autonomous driving.

Individual image quality metrics that focus on a particular form of image degradation have the virtue of being readily decipherable but also the drawback of not relating directly to the purpose for which the image is used. We describe here a method for learning the quality of iris images from the output of iris matching algorithms. We extract a large number of image quality features forming a high dimensional feature vector and label each training image with the match score for its corresponding genuine image in the enrolled database. We then train a Random Forest regressor to predict this match score, and in the course of training apply feature selection to dramatically reduce the feature vector dimensionality. A comparison of several alternative methods of feature selection is given. Our method demonstrates that the effects of image quality degradation on match scores can be predicted from image features. The predicted genuine match score serves as a quality metric, enabling filtering of poor quality images before enrollment or identification.

Traditionally, path planning for field robotic systems is performed in Cartesian space: sensor readings are transformed into terrain costs in a (Cartesian) costmap, and a path to the goal is planned in that map. In this paper, we propose a new approach: planning a path for the robot in the image-space of an on-board camera. We apply a learned color- to-cost mapping to transform a raw image into a cost-image, which then undergoes a pseudo-configuration-space transform. We search in the resulting cost-image for a path to the projected goal point in the image. One benefit of our approach is the ability to react to obstacles at ranges well beyond our 3D sensor range - independent testing has confirmed our system has effectively reacted to obstacles at a range of 93 m while our stereo sensor provides reliable data only up to 5 m away. We describe the details of our technique and the results from testing under the DARPA LAGR and UPI programs.

Driving in unknown natural outdoor terrain is a challenge for autonomous ground vehicles. It can be difficult for a robot to discern obstacles and other hazards in its environment, and characteristics of this high cost terrain may change from one environment to another, or even with different lighting conditions. One successful approach to this problem is for a robot to learn from a demonstration by a human operator. In this paper, we describe an approach to calculating terrain costs from Bayesian estimates using feature vectors measured during a short teleoperated training run in similar terrain and conditions. We describe the theory, its implementation on two different robotic systems, and results of several independently conducted field tests.

Stereo matching in unstructured, outdoor environments is often confounded by the complexity of the scenery and thus may yield only sparse disparity maps. Two-dimensional visual imagery, on the other hand, offers dense information about the environment of mobile robots, but is often difficult to exploit. Training a supervised classifier to identify traversable regions within images that generalizes well across a large variety of environments requires a vast corpus of labeled examples. Autonomous learning of the traversable/untraversable distinction indicated by scene appearance is therefore a highly desirable goal of robot vision. We describe here a system for learning this distinction online without the involvement of a human supervisor. The system takes in imagery and range data from a pair of stereo cameras mounted on a small mobile robot and autonomously learns to produce a labeling of scenery. Supervision of the learning process is entirely through information gathered from range data. Two types of boosted weak learners, Nearest Means and naive Bayes, are trained on this autonomously labeled corpus. The resulting classified images provide dense information about the environment which can be used to fill-in regions where stereo cannot find matches or in lieu of stereo to direct robot navigation. This method has been tested across a large array of environment types and can produce very accurate labelings of scene imagery as judged by human experts and compared against purely geometric-based labelings. Because it is online and rapid, it eliminates some of the problems related to color constancy and dynamic environments.