you can build a pedestrian detector first based on feature extraction like Deformable part models (DPM):
http://www.cs.berkeley.edu/~rbg/latent/
and then after that you can use it to build your pedestrian tracker, the naive approach is to do detection every frame and connect close objects/humans but for better performance, you need to think more on how to make use of it for tracking.
We have recently published a new tracking method which is efficient and effective for crowded scenes in a video-surveillance context. It could be very interesting for you:
http://dx.doi.org/10.1016/j.eswa.2015.06.016
We use given features such as GCH, LBP or HOG. Besides, you can see a video about the performance of our method:
https://www.youtube.com/watch?v=Rqk-vFKzAcQ
Good luck with your research!
Article Expert Video-Surveillance System for Real-Time Detection of ...