1. Vehicle Detection in Aerial Surveillance Using Dynamic Bayesian Networks Hsu-Yung Cheng, Member, IEEE, Chih-Chia Weng, and Yi-Ying Chen IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 21, NO. 4, APRIL 2012

2. Goal

3. Introduction These technologies have a variety of applications, such as military,police, and traffic management. Aerial surveillance is more suitable for monitoring fast-moving targets and covers a much larger spatial area.

4. Introduction Cheng and Butler [8] performed color segmentation via mean-shift algorithm and motion analysis via change detection. In [11], the authors proposed a moving- vehicle detection method based on cascade classifiers. Choi and Yang [12] proposed a vehicle detection algorithm using the symmetric property of car shapes.

5. Introduction

6. Background Color Removal quantize the color histogram bins as 16*16*16. Colors corresponding to the first eight highest bins are regarded as background colors and removed from the scene.

7. Feature Extraction: Local Feature Analysis

9. After evaluation, is known. Use the gradient magnitude G(x,y) of each pixel of moment-preserving. Tmax =T,Tmin=0.1*(Gmax-Gmin) for Canny edge detector. Harris detector is for the corners.

10. Feature Extraction: Color Transform and Color Classification In [16],they proposed a color domain (u,v) instead of (R,G,B) to separate vehicle and non- vehicle pixels clearily. Use n*m as a block to train SVM model to classify color.

11. Feature Extraction: Color Transform and Color Classification

12. Feature Extraction We extract five types of features, S,C,E,A and Z for the pixel. A=L/W Z=blue counts at left

13. Dynamic Bayesian Network Use some videos to train the probabilities with people marked ground truth. Vt indicates if a pixel belongs to a vehicle. P(Vt|St) is defined as the probability that a pixel belongs to a vehicle pixel at time slice given observation St at time Instance t.

14. Experimental results