Process and Implementation: Optical Flow

2395 Words10 Pages
Process and Implementation

Optical Flow

First, I needed to investigate methods to compute optical flow. From the description in the paper, it seemed as if they were using the Lucas-Kanade method. This method assumes that optical flow is constant in a small window. It is a highly used method of computing optical flow and is found in several libraries. The library I utilized was created by Piotr Dollar, from the University of California, San Diego (1). This toolbox also implements a cross-correlation method of computing optical flow, as well as the Horn & Schunck method. I tested all three of these methods and determined that the Lucas-Kanade method was the best choice for this application.

In class, there was a suggestion to investigate a hierarchical version of the Lucas-Kanade method to more accurately calculate optical flow, since it is an integral part of this paper. I tested a toolbox (2) that was written by Sohaib Khan, who was a student of Mubarak Shah at the University of Central Florida. This code produced results that were much more highly susceptible to noise, and I determined that the better choice was the original toolbox. These results can be seen in Figure 1.

Figure 1: Lucas-Kanade Optical Flow Method (left) vs. Hierarchical Lucas-Kanade Optical Flow Method (right)

Generalized Social Force for Particle Advection and Calculation of the Interaction Force

The paper presents a method of using particle advection to estimate the flow of objects through a scene. Since crowd scenes are cluttered and difficult to track microscopically, the authors proposed this macroscopic method of overlaying a grid of particles on the scene and advecting them via optical flow to model the movement of the crowd. This...

... middle of paper ...

... is unclear and a mystery to me. I also would be interested in implementing the concept in a way other than through particle advection, as I feel that it loses a lot of data by not evenly covering the scene in each clip.

Works Cited

1. Dollar, Piotr. Piotr's Matlab Image and Video Toolbox. Piotr Dollar. [Online] February 22, 2010. [Cited: March 17, 2010.] http://vision.ucsd.edu/~pdollar/toolbox/doc/.

2. Khan, Sohaib. Source Code. Mubarak Shah. [Online] [Cited: March 17, 2010.] http://www.cs.ucf.edu/~vision/Code/Optical_Flow/Lucas%20Kanade.zip.

3. Blei, David. Latent Dirichlet Allocation in C. David Blei. [Online] [Cited: March 17, 2010.] http://www.cs.princeton.edu/~blei/lda-c/index.html.

4. Mochihashi, Daichi. lda, a Latent Dirichlet Allocation package. Daichi Mochihashi. [Online] December 4, 2004. [Cited: March 17, 2010.] http://chasen.org/~daiti-m/dist/lda/.
Open Document