Christian Debrunner

Individual recognition from periodic activity using hidden Markov models

We present a method for recognizing individuals from their walking and running gait. The method is based on Hu moments of the motion segmentation in each frame. Periodicity is detected in such a sequence of feature vectors by minimizing the sum of squared differences, and the individual is recognized from the feature vector sequence using hidden Markov models. Comparisons are made to earlier periodicity detection approaches and to earlier individual recognition approaches. Experiments show the successful recognition of individuals (and their gait) in frontoparallel sequences.

Development and analysis of a real-time human motion tracking system

This paper describes a method for tracking human body motion from multiple views in real-time. The method extracts silhouettes in each view using background subtraction, and then intersects the visual hulls generated by the silhouettes to create a set Of voxels. The voxels then exert attractive forces on a kinematic model of the human body to align the model with the voxels. The linked nature of the model allows tracking of partially occluded limbs. The size parameters of the kinematic model are determined automatically during an initialization phase. The kinematic model also incorporates velocity, joint angle, and self collision limits. The entire system with four cameras runs on a single PC in real-time at 20 frames per second. Experiments are presented comparing the performance of the system on real and synthetic imagery to ground truth data.

Stereo vision in LHD automation

This paper details work in applying stereo vision for the enhancement of safety and productivity in the operation of a load-haul-dump (LHD) vehicle in underground mining. The primary goal of this portion of the research is to provide 3D models of the LHDs environment. Availability of these models facilitates performance of automated or teleoperated loading tasks and enhances safety through identification and location of humans in the path of the vehicle. Generation of an accurate 3D model of the immediate surroundings of the LHD is accomplished through processing of stereo visual imagery. Stereo video is acquired using a pair of digital cameras mounted above the cab of the LHD. The video data is processed into a dense depth map plus confidence information. These depths and the stereo rig calibration data are then used to construct a 3D surface model. We demonstrate useful models obtained under both well-illuminated and low-light conditions.

Tomographic reconstruction from an uncontrolled sensor trajectory

For many medical procedures 3D bone models are built from computed tomography (CT) or magnetic resonance imaging (MRI) data, both of which are expensive and time consuming, and unavailable in most operating rooms. We propose 3D tomographic reconstruction from fluoroscopy as an alternative to CT and MRI. Although there are fluoroscopy machines that can perform 3D reconstruction, they must be instrumented and constrained to gather images along a pre-defined path. We present a methodology for 3D reconstruction from imagery collected along an arbitrary path on standard fluoroscopy machines. Using metal beads on the scanned object as fiducial markers, we can recover the path of the X-ray from the image data and apply cone-beam tomographic reconstruction methods. In this paper we describe our cone-beam tomographic reconstruction algorithm and demonstrate reconstruction of a phantom and a cadaver knee from data collected on a standard fluoroscopy unit.

Monte Carlo Visual Tracking Using Color Histograms and a Spatially Weighted Oriented Hausdorff Measure

Color-based and edge-based trackers based on sequential Monte Carlo filters have been shown to be robust and versatile for a modest computational cost. However, background features with characteristics similar to the tracked object can distract them. Robustness can be further improved through the integration of multiple features such that a failure in one feature will not cause the tracker to fail. We present a new method of integrating a shape and a color feature such that even if only a single feature provides correct results, the feature tracker can track correctly. We also introduce a new Hausdorff-based shape similarity metric that we call the spatially weighted oriented Hausdorff similarity measure (SWOHSM). The approach is shown to be robust on both face tracking and automobile tracking applications.

A 2D Fourier Approach to Deformable Model Segmentation of 3D Medical Images

Anatomical shapes present a unique problem in terms of accurate representation and medical image segmentation. Three-dimensional statistical shape models have been extensively researched as a means of autonomously segmenting and representing models. We present a segmentation method driven by a statistical shape model based on a priori shape information from manually segmented training image sets. Our model is comprised of a stack of two-dimensional Fourier descriptors computed from the perimeters of the segmented training image sets after a transformation into a canonical coordinate frame. We apply our shape model to the segmentation of CT and MRI images of the distal femur via an original iterative method based on active contours. The results from the application of our novel method demonstrate its ability to accurately capture anatomical shape variations and guide segmentation. Our quantitative results are unique in that most similar previous work presents only qualitative results.

3D model generation using unconstrained motion of a hand-held video camera

We have developed a shape and structure capture system which constructs accurate, realistic 3D models from video imagery taken with a single freely moving handheld camera. Using an inexpensive off the shelf acquisition system such as a hand-held video camera, we demonstrate the feasibility of fast and accurate generation of these 3D models at a very low cost. In our approach the operator freely moves the camera within some very simple constraints. Our process identifies and tracks high interest image features and computes the relative pose of the camera based on those tracks. Using a RANSAC-like approach we solve for the camera pose and 3D structure based on a homography or essential matrix. Once we have the pose for many frames in the sequence we perform correlation-based stereo to obtain dense point clouds. After these point clouds are computed we integrate them into an octree. By replacing the points in a particular cell with statistics representing the point distribution we can efficiently store the computed model. While being efficient, the integration technique also enables filtering based on occupancy counts which eliminates many stereo outliers and results in an aesthetic viewable 3D model. In this paper we describe our approach in detail as well as show reconstructed results of a synthetic room, an empty room, a lightly furnished room, and an experimental vehicle.

Autonomous vehicle video aided navigation – coupling INS and video approaches

As autonomous vehicle systems become more prevalent, their navigation capabilities become increasingly critical. Currently most systems rely on a combined GPS/INS solution for vehicle pose computation, while some systems use a video-based approach. One problem with a GPS/INS approach is the possible loss of GPS data, especially in urban environments. Using only INS in this case causes significant drift in the computed pose. The video-based approach is not always reliable due to its heavy dependence on image texture. Our approach to autonomous vehicle navigation exploits the best of both of these by coupling an outlier-robust video-based solution with INS when GPS is unavailable. This allows accurate computation of the systems current pose in these situations. In this paper we describe our system design and provide an analysis of its performance, using simulated data with a range of different noise levels.

Fourier Descritpor-Based Deformable Models for Segmentation of the Distal Femur in CT

Anatomical shapes present a unique problem in terms of accurate representation and medical image segmentation. Three-dimensional statistical shape models have been extensively researched as a means of autonomously segmenting and representing models. We present a segmentation method driven by a statistical shape model based on a priori shape information from manually segmented training image sets. Our model is comprised of a stack of two-dimensional Fourier descriptors computed from the perimeters of the segmented training image sets after a transformation into a canonical coordinate frame. Our segmentation process alternates between a local active contour process and a projection onto a global PCA basis of the statistical shape model. We apply our method to the segmentation of CT and MRI images of the distal femur and show quantitatively that it recovers bone shape more accurately from real imagery than a recently published method recovers bone shape from synthetically segmented imagery.

CT from an Unmodified Standard Fluoroscopy Machine Using a Non-reproducible Path

3D reconstruction from image data is required in many medical procedures. Recently, the use of fluoroscopy data to generate these 3D models has been explored. Most existing methods require knowledge of the scanning path either from precise hardware, or pre-calibration procedures. We propose an alternative of obtaining this needed pose information without the need of additional hardware or pre-calibration so that many existing fluoroscopes can be used.