Hello,
I am working on moving object detection using wearable camera.
I wonder if anybody knows how to detect the moving objects when the camera is moving forward?
Motion compensation did will when the camera is rotating, static. But it gives lots of errors when the camera is moving forward/ backward ( big translation).
Any suggestions?
Thanks
According to my comprehension of your question, the dynamic model describing the movement of the camera has to take into account the appropriate model of the error of that motion (forward or backward) for example the problem of wheels glissing . In plus the model of the error when evaluating (in other words; calculating) the depth of the moving object using your wearable camera has to be also correct. I have a doubt about the model of the caculated errors in such cases.
This is a great computer vision question. In addition to @Abedallatif Baba's interesting answer, there is a bit more to add.
There are a couple ways to detect moving objects with a wearable camera such as a fisheye camera used by bikers.
real-time: triangulate each frame and measure the persistence of selected triangles relative to their location in each successive frame. Microprocessors such as the raspberry pi are fast enough so that triangulation and triangle measurement can be done in a few milliseconds.
off-line: repeat real-time steps off-line for each of the frames in a captured video.
@ Abedallatif Baba, thank you so much for your contribution. I found some difficulties calculating camera motion model since it moves in the same plane with the moving objects ( no rigid points in the background because it is already dynamic)!
vSLAM and visual odometry algorithms calculate the real pose of the moving camera but it has to refer to some static key-features, which is not the case in my scenario.
Am I right?
Many thanks again...
@ James F Peters , thanks for your answer.
I divided each frame into rectangular areas and calculated the motion for each sub-rectangle alone. Do you think triangulation would give better results in this case?
Ola
@ Ola Younis: vSLAM and visual odometry algorithms calculate the real pose of the moving camera but it has to refer to some static key-features, which is not the case in my scenario.
Am I right?
In my opinion you can use the "bearing angle" as a global reference of your moving camera and dynamic objects.
@ Ola Younis : I divided each frame into rectangular areas and calculated the motion for each sub-rectangle alone. Do you think triangulation would give better results in this case?
Your approach is very good from a geometry perspective. However, with the formation of rectangles, it is necessary to deal with 4 vertices, namely the corners of the rectangular areas.
With triangulation, there is a reduction in spatial complexity, since it is only necessary to deal with 3 vertices for each triangular region. In addition, each triangle has a barycenter (intersection of the median lines drawn from each vertex to the midpoint of the opposite side) that are helpful in measuring subregions. Add to that the presence of nerve complexes (collections of triangles that have a common vertex) and the machinery that comes from topology, and triangulation looms as a very powerful alternative to rectangulation of image frames.
It was Paul Alexandroff who called attention to the importance of barcentric subdivision. See page 37 in
https://pdfs.semanticscholar.org/8d4e/b721368d4e07b7822908b491ca415c35d1e6.pdf
- Badges
- Science topic