Notes

Outline

Non-parametric Model for Background Subtraction Ahmed Elgammal    David Harwood     Larry Davis Computer Vision Laboratory University of Maryland, College Park MD, 20742, USA

Objectives Robust detection of moving targets in complex outdoor situations from a static camera. High sensitivity detection of real targets. Low false alarm rates. Adaptation to changes in the scene: fast or slow. Work with gray level / color imagery. Suppress shadows from detection. Applications : General outdoor video surveillance systems

Difficulties with Outdoor Scenes Background is not completely static : Tree branches & Bushes movement depends on the wind. Change in lightning conditions: slow or fast Pixel intensity varies significantly over time. One pixel can be image of the sky at one frame, tree leaf at another frame, tree branch on a third frame and some mixture subsequently.

Motivations Intensity distribution is changing dramatically over short periods of time. Modeling intensity variation over long period leads to wide distributions which results in poor detection. Using more “Short-term” distributions will allow better detection sensitivity.

Basic Background Model Capture very recent history about the scene. Continuously updating this history to capture fast changes. Pixel Model : N intensity samples x1,x2,…,xN  taken over time window W. Estimate the probability that a new observed intensity comes from the same distribution using kernel         :

Basic Background Model We use Normal kernel function,               . Σ represents the kernel function bandwidth. Threshold the estimated probability to obtain the foreground.

Kernel Width Estimation Adaptively estimate suitable kernel function bandwidth for each pixel and for each color channel Objective: Measure variation in pixel intensity when the pixel is a projection of the same object. How: Use median, m, of absolute deviation between consecutive intensity values (in time)             to estimate kernel bandwidth.

Suppression of False Detection Suppress Detected pixels that are likely to be displaced from a nearby point (High Pixel Displacement Probability) as a result of: Movements in the background. Camera displacement. Pixel Displacement Probability : Maximum probability that the observed intensity value       belongs to the background distribution of some point in the neighborhood         . Constraint: Component displacement probability                    : The whole detected foreground object (connected component) must have moved from a nearby location.

Suppression of False Detection

Suppression of False Detection

Updating The Background Objective: keep recent, representative samples of pixel intensity. Sample new intensity values periodically. Sample pairs of intensity values (Consecutive frames) Discard old samples (First-in First-out) Select a new samples randomly from each time interval. Update issues : How fast to update ? Where to update ?

Updating The Background How fast to update ? Fast update : what about targets ? Slow update : Wider distributions ! Where to update ? Blind update: update the model for all pixels. Selective update: update only for pixel classified as background.

Updating The Background

Shadow Suppression Use Chromaticity coordinates (                                        ) for probability estimation. Use lightness variable ( s=R+G+B ) to constraint pixel history to “relevant” samples only. Apply probability estimation in the (r,g) space using the constrained samples.

Evaluation Compare results to explicit mixture of Gaussian model. Measure the sensitivity to detect synthetic moving target with low contrast against the background. Exp I: How the False negative rate is affected by target presence in the scene. Exp II: Detection rate for low contrast targets without updating the model (target has no effect on the model)

Evaluation Synthetic Target: moving disk of radius 10 pixels with intensity            . Adjust model parameters to achieve 2% false positive rates.

Results

Results

Implementation Issues Use Pre-calculated lookup tables for the kernel functions. Evaluate partial probability estimations only: Stop kernel calculation when the probability surpasses the required threshold (most image pixels are background pixels.) The implementation of the approach runs at 15-20 frames per second on a 400 MHz Pentium processors for 320x240 gray scale image frames and a background model of 50 sample/pixel.