Roboust Human detection and tracking in a crowded area

Question

I am working on an application where I need to detect and track people in a crowded indoor area(like a mall). Right now I am using the OpenCV Background Subtraction class(MOG2) to detect blobs and a Kalman filter and Hungarian Algorithm for tracking(based on this video).

The issues I'm having are:

1. The blobs merging together when two people come close to each other
2. Parts of the person not getting detected which leads to false and multiple detections on a person
3. The background subtraction itself results in too many false detections.

I would like to know your suggestions to improve this and any solutions to fix these problems? Is there an alternate way to detect humans?I am not using HOG because I didn't get detections unless the entire body of the person was in the frame, and it resulted in false detections as well.

Thanks in advance!

BTW, I'm using OpenCV 3.1,C++

edit:

This what I mean by false detections with HOG:

Answer 1

Take a look of the work done by Rodrigo Benenson.

How Far are We from Solving Pedestrian Detection?

Ten Years of pedestrian detection

It's a really good starting point for pedestrian detection and understand what are the different approaches that has been used in the last decade.