Christopher Paulson

Wavelet-based image registration

Image registration is a fundamental enabling technology in computer vision. Developing an accurate image registration algorithm will significantly improve the techniques for computer vision problems such as tracking, fusion, change detection, autonomous navigation. In this paper, our goal is to develop an algorithm that is robust, automatic, can perform multi-modality registration, reduces the Root Mean Square Error (RMSE) below 4, increases the Peak Signal to Noise Ratio (PSNR) above 34, and uses the wavelet transformation. The preliminary results show that the algorithm is able to achieve a PSNR of approximately 36.7 and RMSE of approximately 3.7. This paper provides a comprehensive discussion of wavelet-based registration algorithm for Remote Sensing applications.

Tracking of Multiple Objects under Partial Occlusion

The goal of multiple object tracking is to find the trajectory of the target objects through a number of frames from an image sequence. Generally, multi-object tracking is a challenging problem due to illumination variation, object occlusion, abrupt object motion and camera motion. In this paper, we propose a multi-object tracking scheme based on a new weighted Kanade-Lucas-Tomasi (KLT) tracker. The original KLT tracking algorithm tracks global feature points instead of a target object, and the features can hardly be tracked through a long sequence because some features may easily get lost after multiple frames. Our tracking method consists of three steps: the first step is to detect moving objects; the second step is to track the features within the moving object mask, where we use a consistency weighted function; and the last step is to identify the trajectory of the object. With an appropriately chosen weighting function, we are able to identify the trajectories of moving objects with high accuracy. In addition, our scheme is able to handle partial object occlusion.

3D dense reconstruction from 2D video sequence via 3D geometric segmentation

3D reconstruction is a major problem in computer vision. This paper considers the problem of reconstructing 3D structures, given a 2D video sequence. This problem is challenging since it is difficult to identify the trajectory of each object point/pixel over time. Traditional stereo 3D reconstruction methods and volumetric 3D reconstruction methods suffer from the blank wall problem, and the estimated dense depth map is not smooth, resulting in loss of actual geometric structures such as planes. To retain geometric structures embedded in the 3D scene, this paper proposes a novel surface fitting approach for 3D dense reconstruction. Specifically, we develop an expanded deterministic annealing algorithm to decompose 3D point cloud to multiple geometric structures, and estimate the parameters of each geometric structure. In this paper, we only consider plane structure, but our methodology can be extended to other parametric geometric structures such as spheres, cylinders, and cones. The experimental results show that the new approach is able to segment 3D point cloud into appropriate geometric structures and generate accurate 3D dense depth map.

A rotation-invariant transform for target detection in SAR images

Rotation of targets pose great a challenge for the design of an automatic image-based target detection system. In this paper, we propose a target detection algorithm that is robust to rotation of targets. Our key idea is to use rotation invariant features as the input for the classifier. For an image in Radon transform space, namely R(b,θ), taking the magnitude of 1-D Fourier transform on θ, we get |Fθ{R(b,θ)}|. It was proved that the coefficients of the combined Radon and 1-D Fourier transform, |Fθ{R(b,θ)}| is invariant to rotation of the image. These coefficients are used as the input to a maximum-margin classifier based on I-RELIEF feature weighting technique. Its objective is to maximize the margin between two classes and improve the robustness of the classifier against uncertainties. For each pixel of a sample SAR image, a feature vector can be extracted from a sub image centered at that pixel. Then our classifier decides whether the pixel is target or non-target. This produces a binary-valued image. We further improve the detection performance by connectivity analysis, image differencing and diversity combining. We evaluate the performance of our proposed algorithm, using the data set collected by Swedish CARABAS-II systems, and the experimental results show that our proposed algorithm achieves superior performance over the benchmark algorithm.

Feature phenomenology and feature extraction of civilian vehicles from SAR images

Being able to recognize one object from another is vital research to our society because it can save lives, improve national security, and improve existing technology such as object avoidance, tracking, etc. In this research we are trying to classify Synthetic Aperture Radar (SAR) images of vehicles from one another no matter if the vehicle is rotated or occluded. The dataset that is being used for this research is the Commercial Vehicle (CV) Data Domes obtained fromWright Patterson Air Force Base (WPAFB). To accomplish this task we used Local Feature Extraction (LFE) to extract the features and then K-nearest neighbor (KNN) was used to classify the vehicles. Overall this method performed well in that the algorithm was able to correctly identify the vehicles 97.6% to 100% accuracy. Currently the algorithm can not handle translation, so the next step of this research is to be able to use the glint information to register the vehicles to a desired location and then perform our algorithm which we believe that registering the image would have a significant improvement to the current results.

Depth-Based Image Registration

Image registration is a fundamental task in computer vision because it can significantly contribute to high-level computer vision and benefit numerous practical applications. Though a lot of image registration techniques exist in literature, there is still a significant amount of research to be conducted because there are a lot of issues that need to be solved such as the parallax problem. The traditional image registration algorithms suffer from the parallax problem due to their underling assumption that the scene can be regarded approximately planar which is not satisfied in the case of large depth variation in the images with high-rise objects. With regard to the the parallax problem, a new strategy is proposed by leveraging the depth information via 3D reconstruction. One novel idea is to recover the depth in the image region with high-rise objects to build accurate transform function for image registration. Our method mitigates the parallax problem and can achieve robust registration results, which is validated by our experiments. Our algorithm is attractive to numerous practical applications.

A target detection scheme for VHF SAR ground surveillance

Detection of targets concealed in foliage is a challenging problem and is critical for ground surveillance. To detect foliage-concealed targets, we need to address two major challenges, namely, 1) how to remotely acquire information that contains important features of foliage-concealed targets, and 2) how to distinguish targets from background and clutter. Synthetic aperture radar operated in low VHF-band has shown very good penetration capability in the forest environment, and hence the first problem can be satisfactorily addressed. The second problem is the focus of this paper. Existing detection schemes can achieve good detection performance but at the cost of high false alarm rate. To address the limitation of the existing schemes, in this paper, we develop a target detection algorithm based on a supervised learning technique that maximizes the margin between two classes, i.e., the target class and the non-target class. Specifically, our target detection algorithm consists of 1) image differencing, 2) maximum-margin classifier, and 3) diversity combining. The image differencing is to enhance and highlight the targets so that the targets are more distinguishable from the background. The maximum-margin classifier is based on a recently developed feature weighting technique called I-RELIEF; the objective of the maximum-margin classifier is to achieve robustness against uncertainties and clutter. The diversity combining utilizes multiple images to further improve the performance of detection, and hence it is a type of multi-pass change detection. We evaluate the performance of our proposed detection algorithm, using the SAR image data collected by Swedish CARABAS-II systems which operates at low VHF-band around 20-90 MHz. The experimental results demonstrate superior performance of our algorithm, compared to the benchmark algorithm associated with the CARABAS-II SAR image data. For example, for the same level of target detection probability, our algorithm only produces 11 false alarms while the benchmark algorithm produces 86 false alarms.

Using glint to perform geometric signature prediction and pose estimation

We consider two problems in this paper. The rst problem is to construct a dictionary of elements without using synthetic data or a subset of the data collection; the second problem is to estimate the orientation of the vehicle, independent of the elevation angle. These problems are important to the SAR community because it will alleviate the cost to create the dictionary and reduce the number of elements in the dictionary needed for classication. In order to accomplish these tasks, we utilize the glint phenomenology, which is usually viewed as a hindrance in most algorithms but is valuable information in our research. One way to capitalize on the glint information is to predict the location of the int by using geometry of the single and double bounce phenomenology. After qualitative examination of the results, we were able to deduce that the geometry information was sucient for accurately predicting the location of the glint. Another way that we exploited the glint characteristics was by using it to extract the angle feature which we will use to do the pose estimation. Using this technique we were able to predict the cardinal heading of the vehicle within ±2° with 96:6% having 0° error. Now this research will have an impact on the classication of SAR images because the geometric prediction will reduce the cost and time to develop and maintain the database for SAR ATR systems and the pose estimation will reduce the computational time and improve accuracy of vehicle classication.