Traffic intersections are critical zones where vehicle and pedestrian interactions significantly impact road safety. This study presents a novel graph-based approach to model and analyze intersection traffic dynamics, leveraging Graph Neural Networks (GNNs) for risk assessment. By representing traffic participants and road infrastructure as a structured graph, we capture spatial-temporal relationships that influence crash likelihood. Using real-world intersection video data, we construct semantic scene graphs to encode actor interactions and road topology, enabling a data-driven understanding of risk factors. Two GNN models, TransformerConv and GINEConv, are employed to assess safety risks, where TransformerConv captures dynamic interactions through adaptive attention weighting, and GINEConv models structured dependencies within the intersection network. Our findings demonstrate that this framework can effectively classify high-risk scenarios, threat assessment of each actors (node), characterize their interaction (edge), and provide near real time safety analysis with 79.8% accuracy. This provides a scalable method for proactive intersection safety monitoring.

Recent work on neural scaling laws demonstrates that model performance scales predictably with compute budget, model size, and dataset size. In this work, we develop scaling laws based on problem complexity. We analyze two fundamental complexity measures: solution space size and representation space size. Using the Traveling Salesman Problem (TSP) as a case study, we show that combinatorial optimization promotes smooth cost trends, and therefore meaningful scaling laws can be obtained even in the absence of an interpretable loss. We then show that suboptimality grows predictably for fixed-size models when scaling the number of TSP nodes or spatial dimensions, independent of whether the model was trained with reinforcement learning or supervised fine-tuning on a static dataset. We conclude with an analogy to problem complexity scaling in local search, showing that a much simpler gradient descent of the cost landscape produces similar trends.

This paper investigates the integration of Large Vision Language Models (LVLMs) with multi-sensor information, including visual and localization data from cameras and LiDAR data to a holistic understanding of traffic videos. Traffic scene understanding is a challenging problem. With complex interaction between the road actors, infrastructure, and traffic rules, it is often difficult to answer questions related to road safety, pedestrian safety, safe maneuvering characteristics, and human factors. Typical processes use a single task-oriented neural network model and combine them through semantic and symbolic reasoning. These processes often suffer from reasoning bias and incompleteness. In recent years, LVLMs have opened new avenues to perceive spatiotemporal information. These models can leverage the large knowledge base from the world and summarize spatiotemporal information effectively. The interactive nature of most of these systems allows humans to directly interact in a visual question-answering mode.In this paper, we have extensively tested the capabilities of such LVLMs to answer key transportation research questions from videos captured through front cameras. We have curated an extensive set of multiple-choice questions to evaluate the performance of these LVLMs. Our results show that LVLMs have abilities to understand various transportation-related aspects to a great extent. Furthermore, we have shown that the addition of supplementary modalities to the VQA settings helps improve the performance of LVLMs. With the addition of 3D trajectories of surrounding objects with the 2D video frames, we observed a significant increase in MCQ performance related to vehicle-to-vehicle interaction tasks. The resources for this paper can be found at https://github.com/sandeshrjain/lvlm-scene

Driver attention is one of the major factors for roadway safety. This paper presents a comprehensive analysis of context-aware driver gaze behavior using large-scale naturalistic data. This paper first demonstrates a Point of Gaze (PoG) estimation system that provides an understanding of the interactions between a driver and the driving environment outside the vehicle. The driver’s PoG positions were mapped to roadway actors that were detected using a road-facing camera, and a driver-facing camera detected the driver’s head and eye motions. This paper presents a new algorithm that can effectively predict gaze fixation to outside objects. The temporal fixation patterns of drivers are studied through a Yarbus-style analysis of PoG. Correlations between gaze direction, salient objects, and the driving scenario are described and analyzed. These correlations have the potential to reveal how drivers behave in various traffic situations and may have the predictive power to aid automated safety improvement systems. Our code is available at: https://github.com/VTTI/gaze-fixation-and-object-saliency and https://github.com/VTTI/Event-Correlation

Distracted driving is one of the leading causes of road accidents. With the recent introduction of advanced driver assistance systems and L2 vehicles, the role of driver attention has gained renewed interest. It is imperative for vehicle manufacturers to develop robust systems that can identify distractions and aid in preventing such accidents in highly automated vehicles. This paper mainly focuses on studying secondary behaviors, and their relative complexity to develop a guide for auto manufacturers. In recent years, a few driver secondary action datasets and deep learning algorithms have been created to address this problem. Despite their success in many domains, Convolutional Neural Network based deep learning methods struggle to fully consider the overall context of an image, and focus on specific image features. We present the use of Video Transformers on two challenging datasets, one of them being a grayscale low-quality dataset. We also demonstrate how the novel concept of a Visual Dictionary can be used to understand the structural components of any secondary behavior. Finally, we validate different components of the visual dictionary by studying the attention modules of the transformer-based model and incorporating explainability in the computer vision model. An activity is decomposed into multiple small actions and attributes and the corresponding attention patches are highlighted in the input frame. Our code is available at github.com/VTTI/driver-secondary-action-recognition

Category-level object pose estimation aims to find 6D object poses of previously unseen object instances from known categories without access to object CAD models. To reduce the huge amount of pose annotations needed for category-level learning, we propose for the first time a self-supervised learning framework to estimate category-level 6D object pose from single 3D point clouds. During training, our method assumes no ground-truth pose annotations, no CAD models, and no multi-view supervision. The key to our method is to disentangle shape and pose through an invariant shape reconstruction module and an equivariant pose estimation module, empowered by SE(3) equivariant point cloud networks. The invariant shape reconstruction module learns to perform aligned reconstructions, yielding a category-level reference frame without using any annotations. In addition, the equivariant pose estimation module achieves category-level pose estimation accuracy that is comparable to some fully supervised methods. Extensive experiments demonstrate the effectiveness of our approach on both complete and partial depth point clouds from the ModelNet40 benchmark, and on real depth point clouds from the NOCS-REAL 275 dataset. The project page with code and visualizations can be found at: dragonlong.github.io/equi-pose.

This paper presents a novel multi-level generative chaotic Recurrent Neural Network (RNN) for image inpainting. This technique utilizes a general framework with multiple chaotic RNN that makes learning the image prior from a single corrupted image more robust and efficient. The proposed network utilizes a randomly-initialized process for parameterization, along with a unique quad-directional encoder structure, chaotic state transition, and adaptive importance for multi-level RNN updating. The efficacy of the approach has been validated through multiple experiments. In spite of a much lower computational load, quantitative comparisons reveal that the proposed approach exceeds the performance of several image-restoration benchmarks.

This paper addresses the task of category-level pose estimation for articulated objects from a single depth image. We present a novel category-level approach that correctly accommodates object instances previously unseen during training. We introduce Articulation-aware Normalized Coordinate Space Hierarchy (ANCSH) - a canonical representation for different articulated objects in a given category. As the key to achieve intra-category generalization, the representation constructs a canonical object space as well as a set of canonical part spaces. The canonical object space normalizes the object orientation, scales and articulations (e.g. joint parameters and states) while each canonical part space further normalizes its part pose and scale. We develop a deep network based on PointNet++ that predicts ANCSH from a single depth point cloud, including part segmentation, normalized coordinates, and joint parameters in the canonical object space. By leveraging the canonicalized joints, we demonstrate: 1) improved performance in part pose and scale estimations using the induced kinematic constraints from joints; 2) high accuracy for joint parameter estimation in camera space.

The current trend in image analysis is to employ automatically detected feature types, such as those obtained using deep-learning techniques. For some applications, however, manually crafted features such as Histogram of Oriented Gradients (HOG) continue to yield better performance in demanding situations. This paper considers both approaches for the problem of facial attribute classification, for images obtained “in the wild.” Attributes of particular interest are eye state (open/closed), mouth state (open/closed), and eyeglasses (present/absent). We present a full face-processing pipeline that employs conventional machine learning techniques, from detection to attribute classification. Experimental results have indicated better performance using RootSIFT with a conventional support-vector machine (SVM) approach, as compared to deep-learning approaches that have been reported in the literature. Our proposed open/closed eye classifier has yielded an accuracy of 99.3% on the CEW dataset, and an accuracy of 98.7% on the ZJU dataset. Similarly, our proposed open/closed mouth classifier has achieved performance similar to deep learning. Also, our proposed presence/absence eyeglasses classifier delivered very good performance, being the best method on LFWA, and second best for the CelebA dataset. The system reported here runs at 30 fps on HD-sized video using a CPU-only implementation.

This paper concerns the assessment of peanut pod maturity through automated visual analysis of the middle layer of the shell, known as the mesocarp. Moisture in the mesocarp decreases with age, resulting in significant changes in color and rigidity. As peanuts mature, the color of the mesocarp gradually changes from white to yellow to dark brown, with progressively darker intensity levels. A technique is presented here in which pod maturity is estimated based on a combination of mesocarp color (visible-light wavelengths) and pod size. For each peanut pod, a voting procedure and nearest-neighbor classification are used to assign 1 of 10 color classes. The color class of each pod, together with the pod's size, are then used to assign 1 of 7 different maturity classes. In the experimental results reported here, precision and recall values for maturity classification were both 66.5%, as compared to classification by a human expert. The proportion of pods in the combined brown-black class, relative to the total sample size, can be used to estimate overall crop maturity. For this task, the automated system identified the number of brown-black peanuts per sample with an accuracy of 92.5%. This approach has strong potential for automatically determining suitable harvest dates for peanut crops.

Flapping flight observed in bats offers a promising model for bio-inspiration of small air vehicles because of their high maneuverability, load carrying capacity, and energy efficiency. However, the flight mechanics of bats is very complex due to the highly articulated wing skeleton and the anisotropic, internally-actuated wing membrane. As a result of these complexities, the shape of bat wings can deform quickly and substantially which causes periodic occlusions and large baseline nonlinear deformations of point trajectories in image space. Tracking these points in image space is difficult because the resolution of the images (720 1280) and the frame rate (120Hz) used in these experiments are substantially lower than those used historically. This paper presents a computational approach that utilizes a novel combination of Laplacian-of-Gaussian (LoG) filtering and optimal threshold segmentation to locate markers in images. Using images from 32 cameras, our technique achieved an average hit rate of 83%, with an average false rate of 12%. Our algorithm is shown to perform better than other techniques, including those based on SIFT or LoG filtering alone. In addition to the improved feature detection algorithm, optical flow based tracking is bootstrapped with a spatially recursive unscented Kalman filter to track the identified points during state estimation. The spatially recursive estimator returns as many or more correct correspondences when compared to the standard unscented Kalman filter.

This paper concerns the detection of human skin in images. Researchers have demonstrated that color information alone is highly effective for skin detection and related segmentation tasks. However, many situations exist for which color information is either not available or not reliable. Examples include images obtained under low-light conditions, legacy grayscale images and videos, and near-infrared images. Furthermore, for human observers it is often the visual texture of skin that is prominent in an image. This paper introduces a novel context-aware framework for skin detection that targets grayscale images. The technique exploits prior geometric knowledge from face detection to localize a set of candidate skin patches in a given image. The system then uses texture-based features (local binary patterns, lacunarity) and statistics derived from grayscale intensities to identify new candidate regions of skin in the image. The system leverages the information resulting from superpixel segmentation to perform region growing. This paper demonstrates feasibility of the approach through experimental results using two standard databases. All the results and the source code will be publicly available 1.

We present a multi-modal feature fusion framework for Kinect-based Facial Expression Recognition (FER). The framework extracts and pre-processes 2D and 3D features separately. The types of 2D and 3D features are selected to maximize the accuracy of the system, with the Histogram of Oriented Gradient (HOG) features for 2D data and statistically selected angles for 3D data giving the best performance. The sets of 2D features and 3D features are reduced and later combined using a novel Dual Kernel Discriminant Analysis (DKDA) approach. Final classification is done using SVMs. The framework is benchmarked on a public Kinect-based FER dataset which includes data for 32 subjects (in both frontal and non-frontal poses and two expression intensities) and 6 basic expressions (plus neutral), namely: happiness, sadness, anger, disgust, fear, and surprise. The framework shows that the proposed combination of 2D and 3D features outperforms simpler existing combinations of 2D and 3D features, as well as systems that use either 2D or 3D features only. The proposed system also outperforms Linear Discriminant Analysis (LDA)-transformed and traditional Kernel Discriminant Analysis (KDA)-transformed systems, with an average accuracy improving of 10%. It also outperforms the state of the art by more than 13% in frontal poses.

Heart rate is a strong indicator of a person's psychophysiological state. For this reason, many applications would benefit from noncontact measurement of heart rate. The paper describes a new procedure for estimating blood volume pulse from a video of a person's face, with an emphasis on real-life scenarios. The approach builds on the algorithm known as Eulerian video magnification, which has shown promise under laboratory conditions, but exhibits problems when attempted in naturalistic situations. In particular, problems arise due to movement by the subject, changing illumination conditions, and low-frame-rate video. This paper describes the procedure that we have developed to address some of these problems, including video rates down to 10 frames per second. The procedure has been tested using videos of indoor subjects, as well as drivers of automobiles in naturalistic situations. The paper also shows analysis and comparison of different stress levels using the extracted heart rate information for a driver on the road.

This paper is concerned with the automatic recognition of pose-varying facial expressions from a low-resolution three-dimensional data sensor, the Kinect. We introduce a novel, comprehensive framework that employs Delaunay triangulation and a pool of Distance Metrics (DM) for feature extraction. Our recognition approach utilizes binary Radial Basis Function (RBF) Support Vector Machines (SVMs) on DM-based feature matrices. Optimal class models are then chosen automatically, and then utilized in the testing stage. We have trained and tested our system using two Kinect-based datasets, with results reaching an average accuracy of 95.1% for non-frontal poses, and more than 98% for frontal poses. Our experimental results show that automatically tuned DMs for each class outperform a fixed DM approach for all classes, especially with non-frontal poses. In addition to the design of the overall framework, this paper describes the effect of training data pruning, providing insights that could contribute to the reduction of training times for very large datasets.

This paper is concerned with the segmentation of noisy point clouds. The ability to partition a set of points into meaningful subsets is of broad interest, spanning such diverse fields as perceptual grouping and remote sensing. We present an approach that is based on projecting the point cloud onto a 2D image grid and applying active contour models (“snakes”) for partitioning point clouds efficiently and effectively. Although active contours were developed for use in image analysis, only a few researchers have considered their application to point-cloud segmentation. Previous systems do not perform well when a high level of noise is present. This paper discusses the heavy dependence of such systems on the initial placement of contours, and we present a novel approach to initializing these systems. Our results demonstrate that good performance is achievable with the approach of geometric active contours (GACs) when appropriately initialized.

Estimating forest canopy heights over large geographical regions is important for quantifying global carbon stocks, which in turn are needed for the study of climate change. Laser altimetry (LiDAR) is increasingly popular as a sensing technique for vegetation height and structure. NASA's ICESat-2, which is scheduled for launch in July 2017, will provide a platform for collecting LiDAR-based canopy height measurements on a global scale over a two-year period. However, unlike the first ICESat mission, which utilized a waveform LiDAR, ICESat-2 will be equipped with a photon-counting LiDAR system. This new sensor is capable of providing increased accuracy in high-slope terrain, but it is also expected to produce large amounts of noise, depending on sun angle and atmospheric conditions. This paper describes recent improvements to a novel two-stage algorithm that estimates ground and canopy surfaces from photon-counting LiDAR, in preparation for the ICESat-2 mission.

Fingerprints continue to serve as a reliable trait for human identification. Feature-based matching techniques, such as those used by Automated Fingerprint Identification Systems (AFIS), have demonstrated remarkable success in minutiae-based matching from good quality prints with relatively large extent. As the image quality degrades and acquired fingerprint area decreases, however, the number of reliable minutiae that can be automatically detected decreases, causing match performance to suffer. This paper presents a novel approach to improving the precision of features that can be extracted from fingerprint images. This is accomplished through improved minutia localization and quality assessment routines that are inspired in part by human visual perception. Initial results have shown an improvement in minutia accuracy for 88.2% of fingerprint minutia sets after applying the proposed localization method. An increase in average quality of true minutiae was found for 98.6% of the fingerprint images when using the proposed quality assessment. The results were obtained using a database of 516 fingerprints with ground truth minutiae.

A watershed transform and a graph partitioning are studied on statistical manifold. Statistical manifold is a 2D Riemannian manifold which is statistically defined by maps that transform a parameter domain onto a set of probability density functions (PDFs). Due to high dimensionality of PDFs, it is hard and computationally expensive to produce segmentation on statistical manifold. In this paper, we propose a method that generates super-pixels using watershed transform. Finding capturing basins on statistical manifold is not straightforward. Here, we create a local distance map using metric tensor defined on statistical manifold. Watershed transform is performed on this local distance map and provides super-pixels that significantly reduce the number of data points and thus make efficient clustering algorithms such as normalized cut (Ncut) feasible to work on. Experimental results show superiority of the proposed method over principal component analysis (PCA) based dimensionality reduction method.

Dimensionality reduction and clustering on statistical manifolds is presented. Statistical manifold [16] is a 2D Riemannian manifold which is statistically defined by maps that transform a parameter domain onto a set of probability density functions. Principal component analysis (PCA) based dimensionality reduction is performed on the manifold, and therefore, estimation of a mean and a variance of the set of probability distributions are needed. First, the probability distributions are transformed by an isometric transform that maps the distributions onto a surface of hyper-sphere. The sphere constructs a Riemannian manifold with a simple geodesic distance measure. Then, a Frechet mean is estimated on the Riemannian manifold to perform the PCA on a tangent plane to the mean. Experimental results show that clustering on the Riemannian space produce more accurate and stable classification than the one on Euclidean space.

This paper introduces a new representation for planar curves. From the well-known Dirichlet problem for a disk, the harmonic function embedded in a circular disk is solely dependent on specified boundary values and can be obtained from Poisson’s integral formula. We derive a discrete version of Poisson’s formula and assess its harmonic properties. Various shape signatures can be used as boundary values, whereas only the corresponding Fourier descriptors are needed for the framework. The proposed approach is similar to a scale space representation but exhibits greater generality by accommodating using any type of shape signature. In addition, it is robust to noise and computationally efficient, and it is guaranteed to have a unique solution. In this paper, we demonstrate that the approach has strong potential for shape representation and matching applications.

This paper presents a new diffusion scheme on statistical manifolds for the detection of texture boundaries. The technique derives from our previous work, in which 2-dimensional Riemannian manifolds were statistically defined by maps that transform a parameter domain onto a set of probability density functions. In the earlier approach, a modified Kullback-Leibler divergence, measuring dissimilarity between two density distributions, was added to the statistical manifolds so that a geometric interpretation of the manifolds becomes possible. Although the previous framework produced good segmentation results, the approach led to offsets in texture boundaries for some situations. This paper introduces a diffusion scheme on statistical manifolds that leads to substantially improved localization accuracy in segmentation of textured images.

A new approach to active contours on statistical manifolds is presented. The statistical manifolds are 2-dimensional Riemannian manifolds that are statistically defined by maps that transform a parameter domain onto a set of probability density functions. In this novel framework, color or texture features are measured at each image point and their statistical characteristics are estimated. This is different from statistical representation of bounded regions. A modified Kullback-Leibler divergence, that measures dissimilarity between two density distributions, is added to the statistical manifolds so that a geometric interpretation of the manifolds becomes possible. With this framework, we can formulate a metric tensor on the statistical manifolds. Then, a geodesic active contour is evolved with the aid of the metric tensor. We show that the statistical manifold framework provides more robust and accurate texture segmentation results.

Future surveillance operations will require easily deployable "microsensors" that are capable of autonomous detection and identification of objects. These devices will operate under severe limitations on energy consumption, to enable battery-powered operation. They will assess sensor inputs locally, transmitting data only after objects of interest have been detected and extracted from sensor data. This paper describes a prototype system that detects and tracks moving objects in image sequences obtained from an infrared video camera. Computation in the system is distributed across an FPGA and a DSP chip. The current system analyzes input images, in search of objects that meet predefined criteria. If these criteria are met, the system extracts a sub-image (or "chip") that contains an object of interest, and then transmits that to a central site for manual review and further analysis.

This paper is concerned with the use of a novel motorized lens to perform segmentation of image sequences. The lens has the effect of introducing small, repeating movements of the camera center, so that objects appear to translate in the image by an amount that depends on distance from the plane of focus. For a stationary scene, optical flow magnitudes and direction are therefore directly related to three-dimensional object distance from the observer. We describe a segmentation-procedure that exploits these controlled observer movements, and we present experimental results that demonstrate the successful extraction of objects at different depths. Potential applications of this approach include video compression, compositing, and passive range sensing.

This paper considers the effect of spatial quantization on several moment invariants. Of particular interest are the affine moment invariants, which have emerged in recent years as a useful tool for image reconstruction, image registration, and recognition of deformed objects. Traditional analysis assumes moments and moment invariants for images that are defined in the continuous domain. In practice, however, the digitization process introduces errors that violate the invariance assumption. This paper presents an analysis of quantization-induced error on (two-dimensional) Hu moment invariants and affine moment invariants, and on invariants derived from (one-dimensional) contour moments. Error bounds are given in several cases.

The paper concerns the recognition of textured objects, such as stained wooden parts, using color images. Many existing color classification systems utilize histogram-based similarity measures to compare an observed image with models from a database. Although the performance of these systems depends heavily on proper quantization of the color space, most quantization methods are based on traditional clustering or thresholding operations. The authors describe a novel approach to color space quantization in which the intersection of meaningful representations results in a partition of the color space. The color descriptions are chosen adaptively, using a set of training images. The resulting partition serves as the domain for histograms of models and of observed images and information-theoretic similarity measures are used to perform recognition. The motivation for this system is to achieve high recognition accuracy in an industrial setting. Laboratory tests have demonstrated a high level of accuracy for this technique, even though the objects of interest exhibit large variations of texture and color.

For indoor wireless communication systems, radio frequency (RF) transceivers need to be placed strategically to achieve optimum communication coverage at the lowest cost. Unfortunately, the coverage region for a transceiver depends heavily on the type of building and on the placement of walls within the building. Traditionally, therefore, transceiver locations have been selected by human experts who rely on experience and heuristics to obtain the optimum (or near-optimum) placement. This paper describes an interactive software system that can be used to assist in transceiver placement. It is intended to be easy to use by individuals who are not experts at wireless communication system design. After the user has selected transceiver locations within a graphical floor plan, the system interprets the floor plan and uses simple path loss models to estimate coverage regions for each transceiver. These regions are highlighted, enabling the user to assess the total coverage. This paper describes the methodology used to compute the coverage regions for multifloored buildings and discusses the effect of interference sources. The resulting system is expected to be useful in the specification of indoor wireless systems.

Presents a new approach for fixing two cameras at a single location in a 3D scene. Fixation requires the detection of binocular disparity for a single point of interest in the scene so that vergence control can reduce that disparity to zero. Most existing systems use area-based matching, since feature-based matching is computationally prohibitive. Unfortunately, the area-based approach does not perform well when confronted with steeply inclined surfaces, occlusions, repeating patterns or featureless image regions. The method presented in this paper utilizes attentional shifts and affine resampling to combat these problems. These are integrated with adaptive window-size control and coarse-to-fine correlation-based searching. The effectiveness of the approach for complex scenes is demonstrated with several stereo image pairs.>

This paper concerns the use of computer vision techniques for distinguishing finished wooden components in a manufacturing environment. The components are kitchen cabinet doors, which are produced in many different shapes and sizes. We have developed a system which can classify doors quickly and reliably as they travel on a conveyor. Two laser sources are used with three video cameras to obtain profile images. This paper describes the system, the feature extraction process, and the classification method. The system exists as a laboratory prototype, and has been successfully tested with a large number of samples. A duplicate system will be installed in a factory in the near future.>

An approach to integrating stereo disparity, camera vergence, and lens focus to exploit their complementary strengths and weaknesses through active control of camera focus and orientations is presented. In addition, the aperture and zoom settings of the cameras are controlled. The result is an active vision system that dynamically and cooperatively interleaves image acquisition with surface estimation. A dense composite map of a single contiguous surface is synthesized by automatically scanning the surface and combining estimates of adjacent, local surface patches. This problem is formulated as one of minimizing a pair of objective functions. The first such function is concerned with the selection of a target for fixation. The second objective function guides the surface estimation process in the vicinity of the fixation point. Calibration parameters of the cameras are treated as variables during optimization, thus making camera calibration an integral, flexible component of surface estimation. An implementation of this method is described, and a performance evaluation of the system is presented. An average absolute error of less than 0.15% in estimated depth was achieved for a large surface having a depth of approximately 2 m.>

A method is described for estimating surfaces from stereo images. A single pair of stereo images can yield surface reconstruction for only a small volume of space. Since the scene may be large, images are obtained dynamically using different camera configurations through the control of focus and camera vergence. For smooth objects, this results in the reconstruction of surface patches that are contiguous and overlapping. These sequentially acquired surface maps are merged into a central. composite representation. During camera reconfiguration, unpredictable calibration errors may arise. Therefore, this method permits small changes in calibration parameter values to achieve agreement between adjacent reconstructed patches in the areas of overlap. To integrate the different sources of depth information, and to balance the effects they have on the resulting surface estimate, the surface reconstruction problem is formulated as one of optimization.>

This paper concerns estimation of surface maps for real scenes having a wide field of view and a wide range of depths. Much research has emphasized stereo disparity as a source of depth information. To a lesser extent, camera focus and camera vergence have also been investigated for their utility in depth recovery. We argue that these sources of visual information have mutually complementary strengths and weaknesses, and to obtain surface maps for real scenes these processes must be integrated. Such integration requires active control of camera orientations and imaging parameters to dynamically and cooperatively interleave image acquisition with surface estimation. Accordingly, a global surface map of the visual field is synthesized by systematically scanning the scene, and combining estimates of adjacent, local surface patches, each acquired by an intermediate camera configuration and having a small depth range. We present an algorithm to perform this integration, and describe its implementation on a dynamic stereo-camera imaging system. Experimental results are presented to demonstrate the superior performance of the integrated system over that of each of its components.

According to Webster, cohesion is “the act or process of sticking together tightly.” Here the term represents the underlying forces that drive the formation of classes during inductive learning. This paper considers several numerical and conceptual induction algorithms, and compares their methods of cohesion. While these algorithms represent several different methods, they also exhibit some significant commonalities.