This is the group project of Digital Image Processing, Tsinghua University, 2015 Spring. We have three great partners:

• Zhengrong Wang
• Hsienyu Meng
• Liuyang Zhan

We tried to do some simple pedestrain tracking based on detection. Use it at your own risk.

The only third party library you will need is OpenCV 3.0.0. And please compile it with VS 2013. Other platform should be fine, but I haven't tested it.

• -b srcVideo tempDiectorary outputImage : Use mean-shift to construct the background.
• -phog srcImage outputImage : Apply HOG detector to a single image.
• -p srcImage bkgImage outputImage : Apply BKG Subtraction + HOG to a single image.
• -vhog srcVideo outputVideo : Apply HOG detector to a video.
• -v srcVideo bkgImage outputVideo : Apply BKG Subtraction + HOG to a video.
• --particle-tracker srcVideo outputVideo : Apply particle filter to track a single target in the video.
• --multiple-tracker srcVideo bkgImage outputVideo : Use detector and particle filter to tracker multiple targets.

For detection part we use simple HOG feature and AdaBoost classifier. We rewrite the baseline program in C++ style and improve the perfomance.

We use particle filter with a online-boosting classifier with RGI feature for each target/pedestrain. We use match matrix and greedy algorithm to handle the data association problem. The main idea is based on this paper.

Here is a simple project website. You can find everything there.

• Documents
• Final Report
• Mean-shift Background Image
• Detector for a video
• Single target tracking result with Gray Haar feature
• Single target tracking result with RGI feature
• Yet another single target tracking result
• Pedestrain Counter result
• Yet another pedestrain counter result