What am I going to do in the next couple of weeks is a fast pedestrian recognition implementation using Histogram of Gradients. (HOG)
HOG , First introduced by Dalal and Triggs1 in 2005 on IEEE International Conference on Computer Vision and Pattern Recognition, is great feature descriptor for human recognition. The basic idea behind it is local object within an image can be described by distribution of intensity gradients or edge direction.
It was very good describing human body and could even extend to other object recognition like vehicles. However, considering the computational cost, practically it was not ideal method as processing time for a 320x240 image would be 1 second.
In the next few years ,several other papers have been published describing faster implementation without sacrificing much accuracy2.
In the other way, without trying to optimize the algorithm itself, people have tried to use GPU to implement it. As CUDA introduced in 2007, heavy tasks could be done on GPU side instead of CPU side.GPU (Graphics Process Units) basically has many 'weak' cores compare to few 'strong' cores on CPU. Then tasks passed to GPU would go 'parallel'.
In this project, a few papers would be primary reference, and OpenCV would be the imaging library to handle CPU side tasks. Also SVM will be used as final classification stage.
Let's see what we'll got in four weeks...
1. Navneet Dalal and Bill Triggs, <Histograms of Oriented Gradients for Human Detection>, 2005
2. Qiang Zhu, Shai Avidan, Mei-Chen Yeh, Kwang-Ting Cheng .<Fast Human Detection Using a Cascade of Histograms of Oriented Gradients>,2006
3. Berkin Bilgic, Berthold K.P. Horn, and Ichiro Masaki<Fast Human Detection with Cascaded Ensembles on the GPU>, 2010
4. Christian Wojek, Gyuri Dorkó, Andre Schulz, Bernt Schiele <Sliding-Windows for Rapid Object-Class Localization: a Parallel Technique>
5. Victor Adrian Prisacariu, Ian Reid, <fastHOG - a real-time GPU implementation of HOG> 2009