kevin_lomax's profile - activity

What exactly do you mean by "identify changes in images"? Do you want to detect new objects that enter the scene or understand the changing weather conditions? The answer could change depend on what your goal is.

For more robustness against weather changes, you can try deep neural networks (esp. convolutional neural nets) provided that you have enough training images. But depending on the actual images, the solution could also be much easier. What are you trying to achieve with background subtraction?

I'm trying out different trackers implemented in OpenCV 3.2 (e.g. MIL, TLD, KCF) on videos where the size of the tracking object changes in 2D plane due to perspective distortion. More precisely, the tracking object is a vehicle which starts moving very far from the camera, at one end of a road, and approaches to the camera (hence becomes bigger) as the time passes.

However, I realized that except for TLD, all the other trackers keep the size of the bounding box (i.e. ROI) same although the size of the tracked object changes dramatically. Is that the intended behavior of these trackers? Is there a parameter I can change to make the bounding box sizes variable? How is this issue typically resolved in practical tracker implementations? I can imagine this can cause some problems when an object starts approaching (moving away from) the camera and appears bigger (smaller) on the image.

Thanks.