Computer Vision

• Introduction
• Projects and Researches

Computer vision, also called machine vision, is a discipline studying the theory, methodology, and algorithms, and system of recognizing, locating objects of interest in 3D world by computer via sensors, such as CCD camera, infrared camera, ultrasound, Satellite imaging, laser, and so on. Contrary to computer graphics, computer vision is challenging research field due to its nature, i.e., it is an inverse process that is trying to understand 3D scenes from imperfect, partial, uncertain, noisy sensored data.

Machine vision has found a wide range of applications in autonomous robotics, missile, inspection, etc. In the past two decades computer vision has been extensively studied, and diverged in different branches such as

• Model-based vision (CAD-based vision)
• Stereo vision (or binocular vision)
• Purposive vision
• Active vision
• Motion analysis and target tracking
• Shape from x (Shading, motion, range data, ...)

• Stabilization of video sequence, department of physics, University Laval, funded by Lockheed Martin Electronic System, Canada.

The aim of the project is to remove the shaking/vibrating effects in image sequence captured by a camera mounted on moving vehicles. The critical part of the project is image registration that registers frames within a specific time interval. Both block matching and feature-based matching can achieve this goal. Usually, block matching is more reliable.
 
• Multi-sensor image fusion, department of physics, University Laval, funded by Lockheed Martin Electronic System, Canada.

The aim of the project is to merge images from different types of sensors to give people more information. Currently, we focus on fusing visible and IR images which are complementary. The vehicles and soldiers in battle field may be occluded by smoke in visible images, but they are visible in IR images. On the other hand, the lands, forests, hills are not well distinguishable in IR images, but they are more perceivable in visible images. The underlying technique for image fusion is image registration. Unlike the previous project, block matching is no longer suitable as the two images to be matched are of different properties. Here, we utilize Hausdorff matching method based on extracted structural information.
 
• Knowledge-guided 3D Reconstruction of Vasculature from Limited Projections, Harbin University of Science and Technology, funded by NSF of China.

As Radon pointed out, 3D reconstruction of objects is impossible unless we have a complete angle of projections. Image reconstruction in CT is carried out by back-projection method which takes dense-angle projections. However, it is possible to reconstruct 3D description of an object if it is sparsely space-occupied and the general knowledge about its formation is known. To reconstruct human blood vessel system is just the case. Although no two individuals' vascular systems are exactly the same, human vascular systems are quite similar. The project create the general elastic model of human blood vessel system by a learning process where the disparity among instances are characterized as elasticity. The model serves as knowledge-base guiding the reconstruction procedure.
 
• 3D Reconstruction of Vasculature from A Pair of Angiograms, Victoria University of Wellington, New Zealand, funded by Wellington Medical Research Foundation.

The project studied vision-based method to reconstruct vasculature from a pair of perpendicular Ray projections. Firstly, the centerline of the vessels are tracked in each image plane. Then their 3D skeletons are reconstructed by standard vision method.
 
• 3D reconstruction from range data funded by NLPR

Range data can be generated by range finder, a device composed of a laser projector and a camera. The well known triangle method is used to calculate the depths of points on object surfaces from the camera. 3D reconstruction from range data has been studied by a number of people. However, one difficulty is the so-called "blind spots", which can not be seen from both laser projector and camera. This project focused on how to recover the lost information based on criteria of consistency and symmetry.
 
• Computing Theory of Machine Vision, NLPR, funded by NSF of China.

In this project, we investigated the foundamental methodology and principles in current vision research. It includes camera calibration, projectve transform and epipolar geometry constraint, 3D reconstruction from point, line segment, and conics correspondences, various numerical methods in vision computations, integration of bottom-up and top-down scheme, etc.
 
• Autonomous Vehicle Navigation, NLPR, funded by "863" High-Tech Plan of China.

Our lab designed and manufactured a vehicle with 6-degree freedom of motions. A platform is mounted on the top of the vehicle, and a pair of cameras are fixed on the platform which have 4-degree of freedom. The vehicle is a test bed for most of our vision method.
In this project, the issues we studied include landmard identification, collision-avoidance, moving target tracking, and navigation in pre-modeled enviroment.