Multi person detection and tracking system

Person detection and tracking consists of three steps

1. Person Detection
2. Person Identification
3. Person Tracking

Person Detection (from External Camera)

There are 3 main competing approaches in OpenCV that are available

1. Histogram of Gradients (HOG) - http://lear.inrialpes.fr/people/triggs/pubs/Dalal-cvpr05.pdf - http://docs.opencv.org/modules/gpu/doc/object_detection.html
2. Haar Cascades - http://research.microsoft.com/en-us/um/people/viola/Pubs/Detect/violaJones_IJCV.pdf - http://docs.opencv.org/modules/objdetect/doc/cascade_classification.html
3. Potentially a combination of the 2? http://lear.inrialpes.fr/people/triggs/student/vj/zhu-cvpr06.pdf - No implementation available

Can we use background subtraction to do better?

Comments

• The HOG approach can be tuned fairly well along the precision-recall curve
• The Haar approach seems to be failing miserably. I can't get to tune it such that so that the results are in a good place in the PR space.
• No profile body detector (pre-trained) is available. We'll have to look through datasets available online to collect representative data, or create our own input. Then we'll have to train according to one of the above approaches. This might not be a limiting thing for the moment as we seem to be producing some number of hits with the frontal detector on sideways motion.

Code: ROS Package camera_person_detector in experimental branch of repository