I want to use the method to estimate depth map of video streams. Objects are moving in the video but the scene is statistic (fixed camera). Also, I don't have any information about camera (eg. Internal or External Calibration).
Unfortunately with uncalibrated cameras you can only get the depth map up to scale, i. e. in unknown units. For that you would need to estimate the Fundamental matrix from pairs of matching points, rectify the images, and compute the disparity map. The disparity is inversely proportional to depth. Please see this example in MATLAB: http://www.mathworks.com/help/vision/ug/stereo-image-rectification.html
However, if you want depth in actual world units you would need to get some additional information, such as actual distances between some points in the scene. Alternatively you would need calibrated cameras or a structured light sensor, like the Kinect.
V. Putz, I have to work on the sample videos and i don't have any access to camera or it's parameters. the Video file is the only thing i have.
But i can consider some assumptions on the movements of objects (the scene is static). Like constant speed motion of objects with some limitation.
How can i use motion blur to estimate depth? ( I just want to know relative depth of objects,not their exact depth)
Could you recommend me the method, related articles and implementations?
Thank you in advance
Thank you Dima Lisin for your clear and helpful response.
The only thing i want is the depth map up to scale, relative depth between moving objects in video.
Is it true to compute Fundamental matrix between consecutive frames which just contain moving objects?
You can try to use a single lens prism based stereovision system. You can try to look up papers with the key word "single lens stereovision system" or "prism-based binocular vision.
Ehsan,
If you only have one fixed camera, then what you need to do is structure from motion:
http://en.wikipedia.org/wiki/Structure_from_motion
The problem becomes much easier if your camera is moving. If the camera motion is arbitrary but fixed on an object you can use the Tomasi-Kanade method. On the other hand, if your scene is fixed, you can initially get the depth of the scene (ignoring moving objects in it), by horizontal translation of the camera. Objects will move outward radially from the optical axis at infinity. The magnitude of this movement (magnitude of this vector) should give you relative (not absolute) depths in the scene. You did not mention whether depth of each moving object should be determined (I'm assuming you need it). In that case you can supplement your depth map with flow vectors to estimate motion (however I'm not sure how I'd do it).
http://link.springer.com/article/10.1007/BF00129684
Mr. Mirsadeghi
If you'd like to check, there is some work which was done regarding the usage of Weightless Neural Networks for both stereo and monocular depth perception.
I guess the monocular case on the papers below would apply to the uncalibrated camera case.
Best Regards
Conference Paper VG-RAM WNN approach to monocular depth perception
Conference Paper Stereo matching with VG-RAM Weightless Neural Networks
- Badges
- Science topic