Image feature descriptors composed of a series of binary intensity comparisons yield substantial memory and runtime improvements over conventional descriptors, but are sensitive to viewpoint changes in ways that vary per feature. We propose a method to improve the matching performance of such descriptors by specifically reasoning about the reliability of test results on a feature-by-feature basis. We demonstrate an intuitive method to learn improved descriptor structures for individual features. Further, these learned results can be efficiently applied during matching with little increase in runtime. We provide an evaluation using a standard, ground-truthed, multi-image dataset.

Visual odometry is typically formulated as a descriptor-based image feature tracking problem, followed by outlier rejection and simultaneous estimation of the scene structure and camera motion. We propose a fundamentally different formulation for the stereo case: a multi-scale search over pose to estimate the transformation that best aligns two sparse point clouds in image space. This has three main consequences. First, data association is descriptorless and implicit, supporting the use of features with indistinct appearance, such as edge features. Second, outlier rejection is subsumed by the use of a robust kernel and a joint feature alignment objective. Third, the method is robust to local minima, in contrast to coarse-to-fine or iterative approaches. This paper details the proposed method, which we call Perspective Alignment Search (PAS), integrated into an edge feature visual odometry system, and an evaluation against a LIDAR-based SLAM solution.

We present a method for learning efficient feature detectors based on in-situ evaluation as an alternative to hand-engineered feature detection methods. We demonstrate our in-situ learning approach by developing a feature detector optimized for stereo visual odometry. Our feature detector parameterization is that of a convolutional filter. We show that feature detectors competitive with the best hand-designed alternatives can be learned by random sampling in the space of convolutional filters and we provide a way to bias the search toward regions of the search space that produce effective results. Further, we describe our approach for obtaining the ground-truth data needed by our learning system in real, everyday environments.

Reliable and accurate camera calibration usually requires an expert intuition to reliably constrain all of the parameters in the camera model. Existing toolboxes ask users to capture images of a calibration target in positions of their choosing, after which the maximum-likelihood calibration is computed using all images in a batch optimization. We introduce a new interactive methodology that uses the current calibration state to suggest the position of the target in the next image and to verify that the final model parameters meet the accuracy requirements specified by the user. Suggesting target positions relies on the ability to score candidate suggestions and their effect on the calibration. We describe two methods for scoring target positions: one that computes the stability of the focal length estimates for initializing the calibration, and another that subsequently quantifies the model uncertainty in pixel space. We demonstrate that our resulting system, AprilCal, consistently yields more accurate camera calibrations than standard tools using results from a set of human trials. We also demonstrate that our approach is applicable for a variety of lenses.

We present a hybrid path planner that combines two common methods for robotic planning: a Dijkstra graph search for the minimum distance path through the configuration space and an optimization scheme to iteratively improve grid-based paths. Our formulation is novel because we first commit to the minimum distance path, then explicitly relax the path to maximize the clearance up to a user-specified bound. Notably, this formulation yields more predictable paths than potential field methods which try to trade increases in path length for greater clearance around obstacles. These potential field costs infer a trade off that can yield poor paths when the obstacle map is partially observable and has a finite history. Some approximations are used to ensure efficient planning, but only a small set of additional behaviors were required to ensure safe operation. Our method has been field tested extensively, as it is the main on-robot path planner for our large team of 14 medium-scale autonomous ground robots and entry to the 2010 Multi Autonomous Ground-robotic International Challenge, MAGIC 2010.

We present an efficient graph-theoretic algorithm for segmenting a colored laser point cloud derived from a laser scanner and camera. Segmentation of raw sensor data is a crucial first step for many high level tasks such as object recognition, obstacle avoidance and terrain classification. Our method enables combination of color information from a wide field of view camera with a 3D LIDAR point cloud from an actuated planar laser scanner. We extend previous work on robust camera-only graph-based segmentation to the case where spatial features, such as surface normals, are available. Our combined method produces segmentation results superior to those derived from either cameras or laser-scanners alone. We verify our approach on both indoor and outdoor scenes.