Towards the goal of autonomous obstacle avoidance for mobile robots, we present a method for superpixel labeling using optical flow templates. Optical flow provides a rich source of information that complements image appearance and point clouds in determining traversability. While much past work uses optical flow towards traversability in a heuristic manner, the method we present here instead classifies flow according to several optical flow templates that are specific to the typical environment shape. Our first contribution over prior work in superpixel labeling using optical flow templates is large improvements in accuracy and efficiency by inference directly from spatiotemporal gradients instead of from independently-computed optical flow, and from improved optical flow modeling for obstacles. Our second contribution over the same is extending superpixel labeling methods to arbitrary camera optics without the need to calibrate the camera, by developing and demonstrating a method for learning optical flow templates from unlabeled video. Our experiments demonstrate successful obstacle detection in an outdoor mobile robot dataset.

Presented at the Ninth Conference on 23rd British Machine Vision Conference (BMVC 2012), 3-7 September 2012, Guildford, Surrey, UK.

Most existing structure from motion (SFM) approaches for unordered images cannot handle multiple instances of the same structure in the scene. When image pairs containing different instances are matched based on visual similarity, the pairwise geometric relations as well as the correspondences inferred from such pairs are erroneous, which can lead to catastrophic failures in the reconstruction. In this paper, we investigate the geometric ambiguities caused by the presence of repeated or duplicate structures and show that to disambiguate between multiple hypotheses requires more than pure geometric reasoning. We couple an expectation maximization (EM)-based algorithm that estimates camera poses and identifies the false match-pairs with an efficient sampling method to discover plausible data association hypotheses. The sampling method is informed by geometric and image-based cues. Our algorithm usually recovers the correct data association, even in the presence of large numbers of false pairwise matches.

We present iSAM2, a fully incremental, graph-based version of incremental smoothing and mapping (iSAM). iSAM2 is based on a novel graphical model-based interpretation of incremental sparse matrix factorization methods, afforded by the recently introduced Bayes tree data structure. The original iSAM algorithm incrementally maintains the square root information matrix by applying matrix factorization updates. We analyze the matrix updates as simple editing operations on the Bayes tree and the conditional densities represented by its cliques. Based on that insight, we present a new method to incrementally change the variable ordering which has a large effect on efficiency. The efficiency and accuracy of the new method is based on fluid relinearization, the concept of selectively relinearizing variables as needed. This allows us to obtain a fully incremental algorithm without any need for periodic batch steps. We analyze the properties of the resulting algorithm in detail, and show on various real and simulated datasets that the iSAM2 algorithm compares favorably with other recent mapping algorithms in both quality and efficiency.

This paper deals with estimation of dense optical flow and ego-motion in a generalized imaging system by exploiting probabilistic linear subspace constraints on the flow. We deal with the extended motion of the imaging system through an environment that we assume to have some degree of statistical regularity. For example, in autonomous ground vehicles the structure of the environment around the vehicle is far from arbitrary, and the depth at each pixel is often approximately constant. The subspace constraints hold not only for perspective cameras, but in fact for a very general class of imaging systems, including catadioptric and multiple-view systems. Using minimal assumptions about the imaging system, we learn a probabilistic subspace constraint that captures the statistical regularity of the scene geometry relative to an imaging system. We propose an extension to probabilistic PCA (Tipping and Bishop, 1999) as a way to robustly learn this subspace from recorded imagery, and demonstrate its use in conjunction with a sparse optical flow algorithm. To deal with the sparseness of the input flow, we use a generative model to estimate the subspace using only the observed flow measurements. Additionally, to identify and cope with image regions that violate subspace constraints, such as moving objects, objects that violate the depth regularity, or gross flow estimation errors, we employ a per-pixel Gaussian mixture outlier process. We demonstrate results of finding the optical flow subspaces and employing them to estimate dense flow and to recover camera motion for a variety of imaging systems in several different environments.

Pose estimation of outdoor robots presents some distinct challenges due to the various uncertainties in the robot sensing and action. In particular, global positioning sensors of outdoor robots do not always work perfectly, causing large drift in the location estimate of the robot. To overcome this common problem, we propose a new approach for global localization using place recognition. First, we learn the location of some arbitrary key places using odometry measurements and GPS measurements only at the start and the end of the robot trajectory. In subsequent runs, when the robot perceives a key place, our fixed-lag smoother fuses odometry measurements with the relative location to the key place to improve its pose estimate. Outdoor mobile robot experiments show that place recognition measurements significantly improve the estimate of the smoother in the absence of GPS measurements.

We present and examine a technique for estimating the ego-motion of a mobile robot using memory-based learning and a monocular camera. Unlike other approaches that rely heavily on camera calibration and geometry to compute trajectory, our method learns a mapping from sparse optical flow to platform velocity and turn rate. We also demonstrate an efficient method of computing high-quality sparse optical flow, and techniques for using this sparse optical flow as input to a supervised learning method. We employ a voting scheme of many learners that use subsets of the sparse optical flow to cope with variable dimensionality and reduce the dimensionality of each learner. Finally, we perform experiments in which we examine the learned mapping for visual odometry, investigate the effects of varying the reduced dimensionality of the sparse optical flow state, and quantify the accuracy of two variations of our learner scheme. Our results indicate that our learning scheme estimates monocular visual odometry mainly from points on the ground plane, and reflect to a degree the minimum dimensionality imposed by the problem. In addition, we show that while this memory-based learning method cannot yet estimate ego-motion as accurately as recent geometric methods, it is possible to learn, with no explicit model of camera calibration or scene structure, complicated mappings that take advantage of properties of the camera and the environment.