How to label/track HOG detections in order to improve the detection system?

asked 2017-10-27 03:35:08 -0500

Jas92 gravatar image

Hi everyone!

I'm working with HOG SVM in order to detect people in a video. I actually trained my own hog detector in order to have a flexible training and detection system. But my own hog detector have many false positives. A way to make it perform better is to eliminate the detections which are present in isolated frames assuming that a person will not appear and disappear suddenly.

-Is there a way to label the hog detections in order to filter them?

-Is it necessary in this case to track the detections ?

-What is the best and fastest way to do it? Is using Kalman filter a good choice? Should I just track the bounding box or do some feature matching inside? and what are the OpenCV available tools that may make this task easier for me?

Sorry for this huge amount of questions asked at once, Your help would be appreciated!



edit retag flag offensive close merge delete


you can try my_detector.yml which i created it to test train_HOG.cpp using INRIAPerson dataset

sturkmen gravatar imagesturkmen ( 2017-10-29 16:26:04 -0500 )edit

I'll check it out, Thanks for sharing @sturkmen

Jas92 gravatar imageJas92 ( 2017-10-29 16:40:15 -0500 )edit

you can use it like in train_HOG.cpp

sturkmen gravatar imagesturkmen ( 2017-10-29 17:31:04 -0500 )edit

if you already download my_detector.yml please download it again.

sturkmen gravatar imagesturkmen ( 2017-10-29 18:01:34 -0500 )edit

if you already download my_detector.yml please download it again. you can test it using train_HOG.cpp like

cpp-example-train_HOG -t -tv=../data/vtest.avi

i trained it using train_HOG.cpp like

cpp-example-train_HOG -d -dw=64 -dh=128 -pd=/INRIAPerson/96X160H96/Train/pos -nd=/INRIAPerson/neg
sturkmen gravatar imagesturkmen ( 2017-10-29 18:01:51 -0500 )edit

@sturkmen, it seems that you've done a very interesting work. I do have some other urgent tasks right now but I will test your code for sure (once possible). Thank you!

Jas92 gravatar imageJas92 ( 2017-10-30 07:43:21 -0500 )edit