Region-based Traffic Video Analysis Using Optical Flow

초록

In this paper, we present a method to analyze traffic video using optical flow estimation. The input video stream was obtained by a camera mounted on a car dash board driven in normal outdoor conditions. Our goal is to analyze the driving and/or navigation patterns from traffic video by defining properties using calculated optical flow throughout the whole video. To analyze the image sequences we define features by estimating velocities from optical flows implemented based on Lucas-Kanade algorithm. Since a traffic image sequence tends to have non-changing backgrounds in specific spatial regions (e.g., dashboard, hood, sky, etc.), we divided the input video frame into multiple regions and rank them with different weights in each region. In our experiments, we show the analysis of driving/navigation patterns extracted from traffic stream images.