vehicle detection and tracking with haar features
Hi,
I am trying to detect and track vehicles by using Haar Feature-based Cascade Classifier in OpenCV. However, I got lots of false positives. Is there any way to eliminate false positives?
Suleyman
I think you have to look into the topic called 'hard negative mining' or 'hard negative learning'
I would try more elaborate techniques, such as OpenTLD
OpenTLD is used to track single objects, also you need to define bounding box by hand. However,I need to detect multiple objects automatically.
Like @sy456 said, OpenTLD won't do the trick here, since it is a tracker system. You still need the detector part before switching to a tracker approach to be completely robust against missing detections.
@StevenPuttemans, I checked Beyond Hard Negative Mining (http://www2.isr.uc.pt/~henriques/beyond/index.html). Afaik, it improves the training part. Have you evert tried it?
Euhm, the title says BEYOND hard negative mining ... ofcourse this is not what you want. Basically what you need to do now is using your existing model with tons of false positives, let it run over lets say 5000 large negative images. Cut out all detections and resize them towards the model size, then use them as negatives in your negative.txt file. Make sure they are at the top so that they are explicitly used for training! With each iteration, your performance will increase but so will the training time to get a new model.
Actually, I used cars3.xml that @albertofernandez mentioned without using negative.txt It is trained with old haar training. I think i need to use train cascade with new negative images.
The cars3.xml is trained with negatives. It is impossible to train a model without specifying what isn't an object...