|
Acivs 2010 Advanced Concepts for Intelligent Vision Systems |
|
|
|
|
Photo: Sally Mayman, Courtesy of Tourism New South Wales |
|
|
Dec. 13-16, 2010 Macquarie University, Sydney, Australia |
Acivs 2010 Abstracts
Invited papers
Paper 251: UAV Video Analysis
Video surveillance and monitoring is one of most active areas of research in Computer Vision. The main steps in a video surveillance system include: detection and categorization of objects of interest in video (e.g. people, vehicles), tracking of those objects from frame to frame, and recognition of their activities, behavior and patterns.
In this talk, I will present an overview of our work in airborne video surveillance. First, I will present our method for tracking thousands of objects in high resolution, low frame-rate, and multiple camera aerial videos. The proposed algorithm avoids the pitfalls of global minimization of data association costs and instead maintains multiple local associations for each track. In contrast with 1-1 correspondence constraints of bipartite graph matching and multiple hypotheses tracking algorithms, the proposed method allows representation of object state in terms of many to many data associations per track.
Next, I will present a method for detecting humans in aerial imagery. The method uses three geometric constraints, namely the orientation of ground plane normal, orientation of shadows cast by humans, and relationship between height and shadow length. This information is used to estimate locations of humans in the scene, and then candidates for detected humans are classified using wavelet features and a Support Vector Machine.
Finally, I will present our approach for detecting key term motion patterns like vehicle turning, pedestrian crossing the road, etc. Motion patterns are spatial segment of the image that has a high degree of local similarity of speed as well as flow direction within the segment and otherwise outside. We employ a mixture model representation of salient patterns of optical flow for learning motion patterns from dense optical flow in a hierarchical, unsupervised fashion. Using low level cues of noisy optical flow, K-means is used to initialize a Gaussian mixture model for temporally segmented clips of video. The components of this mixture are then filtered and instances of motion patterns are computed using a simple motion model, by linking components across space and time.
Paper 252: Remote sensing options
A prime use for remote sensing are agricultural measurements. Analysing multi-spectral images from the earth gives feedback of the quality of the soil and vegetation and can also do predictions of crop-failures and insect plagues, leaving time for action.
Various options are possible for remote sensing. From orbiting satellites to monitoring from unmanned airplanes. They differ in the trade-off between viewing angle and resolution and in the number of spectral bands available. In VITO, Belgium, various remote sensing techniques are used and developed to feed our data-center where daily giga-pixel sized images from the earth are being processed. Also, we work on the use of remote sensing systems for disaster monitoring and for calibration procedures of already installed remote sensing devices.
The talk will give a high level overview of the options in remote sensing.
Paper 253: Convex and Discrete optimization techniques in Computer Vision
Paper 254: Image Analysis - is it just applied statistical analysis and approximation theory?
If one looks across the multitude of papers on solving image analysis problems, there are a limited number of central themes and variations in overall methodolgy.
For example, a paradigm stretching back well into the 70's and probably earlier, is: characterize the problem as the solution of an objective function, study the methods for efficiently solving that optimization problem (including, replacing it with an approximate characterization that is easier to solve).
Popular "schools" such as scale-space analysis, wavelet and (more recently) compressed sensing follow a line of attack that chooses a representation that has certain advantages - particularly leading to approximations with "nice" properties. Much, if not all, of machine learning is a form of statistical approximation (and sometimes entwined with non-statistical, such as geometric, approximation).
This talk will take some influential examples from image-processing/analysis and illustrate how they represent variations on treating the problem as an (possibly statistical) approximation problem. In a sense, the challenge boils down to dealing "the right" approximation that allows a solution methodology that is efficient and robust.