Vehicle Detection and Tracking in Surveillance

1
Vehicle Detection and Tracking in Surveillance

• University of Central Florida
• Andrew Miller, Brandyn White, Arslan Basharat,
  Jingen Lieu,
• and Dr. Mubarak Shah

2
Overview

• Code Split
• Knight System ( SVM classification)
• White Knight (PSO and ICD)
• Abstractions, Python Interface

3
Earlier Work

• KNIGHT System
• Object Detection (Background Subtraction)
• Multi-frame Correspondence Tracking
• Classification (Recurrent Motion)
• Shadow Removal

• Yilmaz, A., Javed, O., Shah, M. Object tracking
  A survey. ACM Comput. Surv. (2006)
• Javed, O., Shah, M. Tracking and ob ject
  classification for automated surveillance. In
  The Seventh European Conference on Computer
  Vision. (Denmark, 2002)
• Javed, O., Shafique, K., Shah, M. A hierarchical
  approach to robust background subtraction using
  color and gradient information. In IEEE Workshop
  on Motion and Video Computing. (Orlando, 2002)
• Shafique, K., Shah, M. A noniterative greedy
  algorithm for multiframe point correspondence.
  IEEE Trans. Pattern Anal. Mach. Intell. (2005)

4
Automatic Parameter Tuning

• Prepare ground-truth segmentations
• A particles position represents a parameter
  configuration
• Use swarm equations to update particles
• Social and Cognitive forces

Brandyn White and Mubarak Shah "Automatically
Tuning Background Subtraction Parameters Using
Particle Swarm Optimization", IEEE International
Conference on Multimedia Expo, Beijing, China
2007.
5
Illumination Change Detection

• Problem illumination changes cause long-term
  artifacts
• Linear regression over time (30 frames)
• Correlation coefficient indicates trend of
  illumination change
• Temporarily increase learning rate

6
SVM Classification

• Nine-dimensional feature vector
• Eight edge-orientation histogram bins
• Aspect ratio
• Linear classifier

7
Results

• VACE Core (September 2006)
• Mean MOTA Mean MOTA 51.5 (Unofficial result)
• CLEAR (May 2007)
• (KNIGHT) Mean MOTA 53.3 MOTP 55.9
• (WKNIGHT) Mean MOTA 22.5 MOTP 63.7
• Didnt finish coding White Knight
• Simple tracking and classification
• How much does background subtraction matter?
• The goal of evaluations is to learn how to answer
  this question rapidly and methodically

8
Results MOTA Boxplots

• Statistically Significant?
• High degree of per-sequence variation

Knight
White Knight
9
Results Per-Sequence Comparison?

• Precision vs Accuracy
• Mean Absolute Difference vs Mean Difference?

Red (SVM) Blue (PSO and Illumination)