Entropy will give you the amount of information contained in a picture.Two adjacent pictures in a video stream may have the same amount of information content but with the the subject moved to a different area of the picture. Using entropy alone will not tell you that the subject has moved. A good place to start would be looking at some basic texts on video compression which will give you an understanding of motion vectors between adjacent pictures. There is generally lots of Matlab, C++ and Python code around that you can use to play with this stuff.

It seems to me that a sum, or an average, depending of what you want to measure, of the motion vectors per frame and for all pixels for optical flow, or for all blocks for hybrid block video compression (HEVC, H264, MPEG2), would provide a correct result of total motion for a sequence of frames. But I do not know any specific metric up to now. You can also have a look at these works : http://lcni.uoregon.edu/~dow/Geek_software/Video_motion/Video_motion_estimation.html

http://people.ece.cornell.edu/land/PROJECTS/MotionDamian/

I've been looking in ways to do it. I am trying to find a metric for a dynamic event observed by two imagers. And I want that metric to tell me which one is better. Obviously static imaging evaluation using well established metric like MTF, NPS and DQE give you some information but not the entire picture. SO I am using some RMS approach, you can use the formulas from this SPIE paper I did last year: Evaluation of intracranial aneurysm coil embolization in phantoms and patients using a high-resolution Microangiographic Fluoroscope (MAF).

Link to full article: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3419595/

You may have a look at the MPEG-7 Visual Motion Descriptors.

This is a very tricky question, as you refer to "object" motion. This technically would require a clear segmentation of the objects in videos first. Actually much of the research into innovations in video codecs comes from exactly this point - how to separate the mess of pixels into coherently moving, and hence more compactly describable, entities. So yes, any modern video codec (if you have the source code for it) would give you access to motion regions and their amount of movement as predicted from past data.

Optic flow - as you mentioned - is also useful. Again, applied without segmentation, it will describe the amount of pixel movement in the image - also note, that motion in depth, occlusions, appearances and disappearances of objects, lighting effects, etc. will all result in highly non-linear effects on optic flow.

Most of the research in computer vision nowadays focuses on object-class specific motion analysis, such as facial motion, body motion, articulated motion, crowd motion, or on specific subtypes, such as rigid-motion, etc.

Hi, maybe you can plot a 2D histogram of the optical flow distribution. This can give you a good idea of how the motion is distributed in the video sequence. Also, you can classify the objects as static and dynamic according with the result of the optical flow algorithm, then you can exactly know which pixels are static and dynamic. Then you can compute a mean value over the entire optical flow or just for the moving objects. In any case you have a measure of the mean motion in the video sequence. Try something simple at the beginning, for instance Lucas-Kanade algorithm.

http://www.mathworks.com/matlabcentral/fileexchange/23142-iterative-pyramidal-lk-optical-flow/all_files

Typically, you can take the residual energy between the adjacent frames, and then apply any transform (like DCT or integer transform). After getting the transform coefficients, you can do entropy coding and then compare the amount of motion activity between two videos.

Exact answer I would give is dependent on the application whether the test is real time or offline processing.

Real Time Apps i would looking to Homography and very simple feature matching between frames which can give you an idea of exactly how much each part of the image has moved and by how much.

Offline Apps i would go for the Convolution and full on entropy testing of the image as the other have mentioned above. Which can give you a much better over all idea of the movement in the image but can be more time consuming but obviously can give you a pixel by pixel account of movement.

Good Luck

if the variable illumation does not produce changes in the visualization of the object, I would recommend you the extension of Baddeley metric to grayscale. It is not computationally cheap, though.

Application of Baddeley's distance to dissimilarity measurement between gray scale images- Coquin, Bolon, PRLetters (2001)

https://www.researchgate.net/publication/220644330_Application_of_Baddeley%27s_distance_to_dissimilarity_measurement_between_gray_scale_images?citationList=incoming

It can be used on color images without difficulties.

Article Application of Baddeley's distance to dissimilarity measurem...

Block matching algorithms results in motion vectors which are considered good for calculating motion. As you are not requiring estimating exact block displacement, I would suggest the use of a simple block matching algorithm such as the TSS, DS, or Cross search, which can be followed by simple averaging of MVs.

How about to use as metrics the derivative of the distance variation in time between the center of the screen and the center of the object?

Have a look at this paper: "Orthogonal variant moments features in image analysis" There is a discusion about measuring object motion and a way to calculate it.

José Antonio

Could you please send me the paper: Orthogonal variant moments features in image analysis, because I have no acces to it? Thanks.

Hola Marta, aquí está.

Saludos.

Article Orthogonal variant moments features in image analysis

ḿuchas gracias Antonio!!!

In video compression standards (starting from the ITU-T H.261 to the latest and greatest H.265), motion estimation are measured in pixels, and the result of motion estimation which is encoded as motion vectors are expressed in pixels. The range of motion that can be expressed in motion vectors depend on the compression standard. So if the standard allows for 5 bits. then motion vectors between 0 and 31 pixels are considered. Obviously motion estimation is the most expansive procedure in video compression accounting typically around 40 to 45% of the compression computation.

Another alternative could be to compute trajectories (i.e. path lines) over a finite-time interval in the time-dependent optical flow field and evaluate the geometric properties of those trajectories (e.g. via their arc-length). Roughly spoken, this corresponds to an accumulation of motion vector magnitudes at a specific position over the considered time interval.

There are many off-the-shelf tools or algorithms, particular during MPEG 7 for motion related descriptor.

Mark it

Better to go for motion estimation which means calculation of motion vectors

Euclidean distances of the centriod of the object can be one criteria to estimate the amount of motion. This is simpler.

Hello

Here is the Matlab Code Example for tracking ball, it may be helpful for you. You can modify it according to your requirement.

%Code Example

b=imaqhwinfo %show adaptor

camera_name = char(b.InstalledAdaptors(end))

camera_info = imaqhwinfo(camera_name)

vid= videoinput('winvideo',[1]) %Summary of Video Input Object

a = getselectedsource(vid) %Display Summary for Video Source Object

get(a)

% Set the properties of the video object

set(vid, 'FramesPerTrigger', Inf);

set(vid, 'ReturnedColorspace', 'rgb')

vid.FrameGrabInterval = 5;

%start the video aquisition here

start(vid)

% Set a loop that stop after 100 frames of aquisition

while(vid.FramesAcquired