Question

I’ve got a ConcurrentQueue which is being populated with a stream of images. For simplicity’s sake, assume they’re frames from a webcam.

What I’d ideally like to do is detect

• if there’s any motion at all
• Where the largest (by size not speed) motion is in the frame
• Where the second largest motion is in the frame

Presumably I need to composite a reference image over the last N frames (so that semi-permanent changes are accounted for eg day/night, parked cars, etc…) and then difference the current frame from the reference frame.

The frames have a (minimal) level of noise so exact colour-comparison isn’t a viable option.

I’m sure I’ve seen tutorials on this before (culminating in a “box” around large areas of movement) but I can’t find them now.

Can someone point me at a decent algorithm/tutorial?

Answer 1

If you just want a solution that works, ZoneMinder or Motion are two pieces of software that run under linux using the video4linux interface.

If you need to roll your own for some reason there are a lot of techniques or strategies you can use. You are largely on the right track with what you’ve outlined. You’re missing a few important details though.

1. Since the camera is stationary, keep a record of the last N frames as your “background” image. Average them all.

   http://opencv.willowgarage.com/documentation/cpp/imgproc_motion_analysis_and_object_tracking.html

2. Subtract the background from the current image. What you’re left with we’ll call the foreground.

   http://opencv.willowgarage.com/documentation/cpp/core_operations_on_arrays.html#cv-absdiff

3. Optionally perform dilation or erosion (or both) to remove noise or join nearly connection regions.

   http://opencv.willowgarage.com/documentation/image_filtering.html#dilate

4. Threshold the foreground image to determine what’s important and what’s not.

   http://docs.opencv.org/doc/tutorials/imgproc/threshold/threshold.html

5. Optionally use the findContours function to get a description of what’s “moved”

   http://docs.opencv.org/doc/tutorials/imgproc/shapedescriptors/find_contours/find_contours.html

   Once you have the contours you can also find the bounding rectangles if that’s more what you’re going for.

   http://opencv.willowgarage.com/documentation/python/structural_analysis_and_shape_descriptors.html#boundingrect

This will not be perfect and when debugging or optimizing you have to show output after every step to figure out what’s working right and what isn’t. Spend some time building the infrastructure to make that easier. Once you have source data and most of a working pipeline tuning to get the results you want is quite doable.