Structural learning, a method to estimate the parameters for discrete energy minimization, has been proven to be effective in solving computer vision problems, especially in 3D scene parsing. As the complexity of the models increases, structural learning algorithms turn to approximate inference to retain tractability. Unfortunately, such methods often fail because the approximation can be arbitrarily poor. In this work, we propose a method to overcome this limitation through exploiting the properties of the joint problem of training time inference and learning. With the help of the learning framework, we transform the inapproximable inference problem into a polynomial time solvable one, thereby enabling tractable exact inference while still allowing an arbitrary graph structure and full potential interactions. Our learning algorithm is guaranteed to return a solution with a bounded error to the global optimal within the feasible parameter space. We demonstrate the effectiveness of this method on two point cloud scene parsing datasets. Our approach runs much faster and solves a problem that is intractable for previous, well-known approaches.

This paper studies the problem of GPS-denied unmanned ground vehicle (UGV) localization by matching ground images to a satellite map. We examine the realistic, but particularly challenging problem of navigation in remote areas using maps that may correspond to a different season of the year. The problem is difficult due to the limited UGV sensor horizon, the drastic shift in perspective between ground and aerial views, the absence of discriminative features in the environment due to the remote location, and the high variation in appearance of the satellite map caused by the change in seasons. We present an approach to image matching using semantic information that is invariant to seasonal change. This semantics-based matching is incorporated into a particle filter framework and successful localization of the ground vehicle is demonstrated for satellite maps captured in summer, spring, and winter.

Many learning-based computer vision algorithms perform poorly when faced with examples that are dissimilar to those on which they were trained. Domain adaptation methods attempt to address this problem, but usually assume that the source domain is specified a priori. We propose a two-step approach for situations where more than one source domain is available. The first step uses a small number of labeled examples to choose the source domain most similar to the target domain, while the second step uses traditional domain adaptation methods to further adapt the chosen source domain to the target data. We demonstrate this two-step domain adaptation algorithm in the context of style-independent building component recognition, which suffers from the problem of inter-domain performance degradation. In this case, different building styles represent the domains, and the task is to reverse engineer a new building of unknown style. We evaluate several variants of the two-step method, and experiments show that the proposed approach outperforms existing single-step methods on a dataset of nine building styles. We demonstrate the generality of the approach on a large, multi-domain dataset with 22 product review categories (i.e., Styles) from the natural language processing field.

Applications, such as construction monitoring and planning for renovations, require the accurate recovery of existing conditions of structures. Many types of infrastructure are primarily comprised of arbitrarily-shaped thin structures (e.g., Truss bridges, steel frame buildings under construction, and transmission towers), which existing automatic modeling methods are incapable of handling. To address this issue, this paper presents an approach to automatically recognize and model beams, planes, and joints from a 3D point cloud containing a complex network of thin structures, and to recover their topology. In our approach, each beam is evolved from a seed by matching and aligning the cross section images. This growing algorithm can model beams with arbitrary cross sections. By performing the algorithm on a point connectivity graph, we distinguish beams from joints and improve the algorithm's robustness to closely spaced objects. In parallel, planes and joints are also extracted and modeled. The connectivity graph of these primitives allows for a compact, object-level understanding of the entire structure. We demonstrate the capability and robustness of our approach on both synthetic and real datasets.

With the emergence of relatively low-cost real-time 3D imaging sensors, new applications for suites of 3D sensors are becoming practical. For example, 3D sensors in an industrial robotic work cell can monitor workers' positions to ensure their safety. This paper introduces a simple-to-use method for extrinsic calibration of multiple 3D sensors observing a common workspace. Traditional planar target camera calibration techniques are not well-suited for such situations, because multiple cameras may not observe the same target. Our method uses a hand-held spherical target, which is imaged from various points within the workspace. The algorithm automatically detects the sphere in a sequence of views and simultaneously estimates the sphere centers and extrinsic parameters to align an arbitrary network of 3D sensors. We demonstrate the approach with examples of calibrating heterogeneous collections of 3D cameras and achieve better results than traditional, image-based calibration.

This paper studies the problem of matching images captured from an unmanned ground vehicle (UGV) to those from a satellite or high-flying vehicle. We focus on situations where the UGV navigates in remote areas with few man-made structures. This is a difficult problem due to the drastic change in perspective between the ground and aerial imagery and the lack of environmental features for image comparison. We do not rely on GPS, which may be jammed or uncertain. We propose a two-step approach: (1) the UGV images are warped to obtain a bird's eye view of the ground, and (2) this view is compared to a grid of satellite locations using whole-image descriptors. We analyze the performance of a variety of descriptors for different satellite map sizes and various terrain and environment types. We incorporate the air-ground matching into a particle-filter framework for localization using the best-performing descriptor. The results show that vision-based UGV localization from satellite maps is not only possible, but often provides better position estimates than GPS estimates, enabling us to improve the location estimates of Google Street View.

We evaluate the performance of a widely used tracking-by-detection and data association multi-target tracking pipeline applied to an activity-rich video dataset. In contrast to traditional work on multi-target pedestrian tracking where people are largely assumed to be upright, we use an activity-rich dataset that includes a wide range of body poses derived from actions such as picking up an object, riding a bike, digging with a shovel, and sitting down. For each step of the tracking pipeline, we identify key limitations and offer practical modifications that enable robust multi-target tracking over a range of activities. We show that the use of multiple posture-specific detectors and an appearance-based data association post-processing step can generate non-fragmented trajectories essential for holistic activity understanding.

Current manufacturing practices require complete physical separation between people and active industrial robots. These precautions ensure safety, but are inefficient in terms of time and resources, and place limits on the types of tasks that can be performed. In this paper, we present a real-time, sensor-based approach for ensuring the safety of people in close proximity to robots in an industrial workcell. Our approach fuses data from multiple 3D imaging sensors of different modalities into a volumetric evidence grid and segments the volume into regions corresponding to background, robots, and people. Surrounding each robot is a danger zone that dynamically updates according to the robot's position and trajectory. Similarly, surrounding each person is a dynamically updated safety zone. A collision between danger and safety zones indicates an impending actual collision, and the affected robot is stopped until the problem is resolved. We demonstrate and experimentally evaluate the concept in a prototype industrial workcell augmented with stereo and range cameras.

We have developed software for fully automated tracking of vibrissae (whiskers) in high-speed videos (>500 Hz) of head-fixed, behaving rodents trimmed to a single row of whiskers. Performance was assessed against a manually curated dataset consisting of 1.32 million video frames comprising 4.5 million whisker traces. The current implementation detects whiskers with a recall of 99.998% and identifies individual whiskers with 99.997% accuracy. The average processing rate for these images was 8 Mpx/s/cpu (2.6 GHz Intel Core2, 2 GB RAM). This translates to 35 processed frames per second for a 640 px×352 px video of 4 whiskers. The speed and accuracy achieved enables quantitative behavioral studies where the analysis of millions of video frames is required. We used the software to analyze the evolving whisking strategies as mice learned a whisker-based detection task over the course of 6 days (8148 trials, 25 million frames) and measure the forces at the sensory follicle that most underlie haptic perception.

Evidence grids are a popular representation for fused data from multiple sensors. Previous attempts at back ground subtraction within evidence grids either do so prior to sensor fusion or do so naively, simply ignoring any cells with a high background occupancy probability. A key weakness of these approaches is that they cannot reason about interiors of objects or other unobserved regions. Recognizing and removing solid object interiors is important for any application that must be able to differentiate between occupied and unknown space after background subtraction. In this paper, we propose accessibility analysis as a method for the removal of interior regions. We then present and compare two approaches for performing background subtraction with accessibility analysis in evidence grids. Performance is measured using a 3D evidence grid in a test bed for a sensing system designed for use in safety monitoring of an automated assembly workcell. Within the parameters of the present study, both techniques allow for precise detection of foreground objects while fully removing background objects. Subtraction runs in near real-time, even for large grids.

Semantic 3D models of buildings encode the geometry as well as the identity of key components of a facility, such as walls, floors, and ceilings. Manually constructing such a model is a time-consuming and error-prone process. Our goal is to automate this process using 3D point data from a laser scanner. Our hypothesis is that contextual information is important to reliable performance in unmodified environments, which are often highly cluttered. We use a Conditional Random Field (CRF) model to discover and exploit contextual information, classifying planar patches extracted from the point cloud data. We compare the results of our context-based CRF algorithm with a context-free method based on L2 norm regularized Logistic Regression (RLR). We find that using certain contextual information along with local features leads to better classification results. 1

For an autonomous vehicle, detecting and tracking other vehicles is a critical task. Determining the orientation of a detected vehicle is necessary for assessing whether the vehicle is a potential hazard. If a detected vehicle is moving, the orientation can be inferred from its trajectory, but if the vehicle is stationary, the orientation must be determined directly. In this paper, we focus on vision-based algorithms for determining vehicle orientation of vehicles in images. We train a set of Histogram of Oriented Gradients (HOG) classifiers to recognize different orientations of vehicles detected in imagery. We find that these orientation-specific classifiers perform well, achieving a 88% classification accuracy on a test database of 284 images. We also investigate how combinations of orientation-specific classifiers can be employed to distinguish subsets of orientations, such as driver's side versus passenger's side views. Finally, we compare a vehicle detector formed from orientation-specific classifiers to an orientation-independent classifier and find that, counter-intuitively, the orientation-independent classifier outperforms the set of orientation-specific classifiers.

Noncoherently detected frequency shift keying can be used in fast fading environments where it is difficult to track the channel. In this paper we present two methods to improve the performance of iteratively detected OFDM-MFSK. OFDM-MFSK is a combination of orthogonal frequency division multiplexing (OFDM) and noncoherent M-ary frequency shift keying (MFSK) that was developed as a robust transmission method for fast time varying channels. The first method increases the number of bits per label and therefore leads to an increased data rate, but also to an ambiguous mapping which has to be resolved by the channel decoder. In the second method, a precoder is inserted between the channel coder and the mapper, leading to a system of three serially concatenated codes. To analyze the iterative behavior of the receiver, we will use three dimensional EXIT charts, which allow us to predict the performance. The intention of both methods is to adapt the channel code to the modulation scheme in order to improve the performance of the iterative receiver.

In this paper we present a system for the simultaneous visualization of several parallel executed simulation replications. By aggregating the scenes of multiple similar simulations into one single scene it is possible to make a visual statistical analysis of a set of discrete event simulations as well as to easily compare different system parameterizations. The aim of our system is to enhance the model analysis, verification and validation process in terms of speed and ease. The parallel execution of several simulations of complex models and the visualization of these cannot be done on one computer, thus a parallel approach is necessary.Our system uses a thin-client and multiple processorson a PC-cluster. The rendering and the simulation executionare done on processors of the cluster. The client is usedonly for the visualization of the images transmitted by thecluster and for user interaction.

Measurements of impedance spectra used for non-invasive glucose monitoring are affected by a variety of perturbing factors such as temperature and sweat/moisture fluctuations, changes in perfusion, and body movements. In order to quantify and compensate for these perturbing effects, a multi-sensor approach was suggested. Different sensors are used, measuring signals correlated with blood glucose and perturbing factors, respectively. Here, we investigate how the multiple sensor data can be transformed into meaningful information about changes in the concentration of blood glucose. Linear regression models and variable selection (stepwise for/back-ward and lasso) techniques are used to derive generally valid models allowing for the estimation of blood glucose concentration. We find that over-fitting is best avoided by using a special version of cross-validated prediction error as the model selection criterion. Indeed, the resulting models are reasonably small, plausible, and comprise an additive adjustment for the experimental run.