ICIP 2006, Atlanta, GA

Slide Show

Atlanta Conv. & Vis. Bureau


Technical Program

Paper Detail

Session:Video Object Segmentation and Tracking
Time:Tuesday, October 10, 14:20 - 17:00
Presentation: Poster
Authors: Sascha Cvetkovic; Bosch Security Systems 
 Peter Bakker; Bosch Security Systems 
 Johan Schirris; Bosch Security Systems 
 Peter H.N. de With; University of Technology Eindhoven / LogicaCMG 
Abstract: In surveillance applications, the camera has typically a fixed position and for security reasons, system users are particularly interested in the moving regions in the video signals differing from the (possible) camera motion. Background (BG) subtraction is used to segment moving regions in image sequences, by comparing each new frame to a model of the scene BG. In this paper, we present an efficient non-parametric BG model and a BG-subtraction algorithm that require only the luminance channel of a heavily down-sampled video. The model has three advantages. First, the model can handle complex situations where the BG of the scene is cluttered and not completely static. Second, the model can also extrapolate luminance values to estimate pixel trends and incorporate medium-speed changes, and it adapts quickly to scene changes to facilitate a very sensitive detection of moving targets. Third, it learns objects that are detected as foreground (FG) for a long time and we propose a solution for learning the BG objects starting to move after an initial BG model was created. At the same time, we found a way to reduce memory size significantly. This paper also shows the effect of heavy down-sampling of the input video to QCIF resolution and the influence of a local image enhancement method (CLAHE) that is applied prior to the down-sampling to circumvent sensitive quality deterioration. Due to the heavy down-sampling of the video signal, image noise does not have a normal distribution anymore, but it behaves like quantization noise which deteriorates the operation of common light-change detectors. We present a light-change removal algorithm that uses only the luminance signal and discuss its performance. In cases when numerous light-changes are detected, we see that the use of light-change detection is compulsory in such extreme conditions to cope with false positives. Applying local contrast enhancement prior to down-sampling sometimes increases the sensitivity of the foreground detection and contributes to the overall result especially for small low-contrast foreground objects, which otherwise would be detected as a light-change. Often, the performance is a compromise between merging foreground objects, the amount of false positives and the false negatives. Down-sampling generally improves the performance of light-change detection (less false positives. However there is a limit to that since local contrast can be significantly lowered due to the down-sampling which often can lead to detection of FG objects as light changes or complete miss-detection of FG objects (the FG detection works better on original resolution images). We succeeded in making a system that can still work with a decent performance on such a dataset in real time (25fr/s progressive, CIF/QCIF format, luminance channel only).