My field of work is related to developing interactive applications using computer vision. Most of interactive applications use 3D sensors that are able to map the real distance from each pixel on the field of view to the camera (p.e. kinect), building an accurate depth map, which has a huge amount of possible applications in many fields.
For the past few days I've been studying the state of the art of stereo vision, to see if it can in any way replace 3D sensors, but my investigation is turning out to be frustrating, since I can't find anything that may suit my needs.
So, my question is this: How far advanced is this technology? Are there any stereo vision algorithms that can generate an accurate depth map at 30 frames per second? Or we are just not there yet?
Could you please tell me the implications of the height resolution being 5 to 10 mm? Does it mean that ToF will only recognize height differences of that magnitude? Or it means that each height reading may have a 5 to 10 mm of error? If it is the latter, it would work for the application. If it is the former, it wouldn't work.
Yes, IR projections wouldn't in fact be visible to the human eye, could in fact work to help creating an accurate depth map.
In ToF, the system has to measure the time the photon takes to fly to the object's surface and back to the camera. For a distance of 10 mm this corresponds to about 33 ps (Pikoseconds) which is extremely short and difficult to measure.
This is why typical quantization of height values in a ToF system are around 10 mm; this is the resolution. The error or standard deviation of a height reading is even coarser by a factor of typically 5 to 10; but this is true for all 3D imaging techniques, also for stereo.
Got it, thanks !
What do you make of this link? http://www.ti.com/ww/en/analog/3dtof/index.shtml?DCMP=hpa_contributed_article
I did not know that TI also offers ToF sensors. However, the lateral resolution is not very high. The nice thing is that TI offers all peripheral electronics that is necessary to get a 3D image out of the ToF sensor. But also other suppliers of ToF camera systems normally have embedded solutions and do the 3D processing of FPGA or DSPs. As I have written in my first comment, there are very promissing ToF approaches with 1.3 MPixel sensors @50fps which employ a direct pulsed ToF principle that avoids the distance ambiguity of a modulated IR approach; TI is modulated, odos is pulsed.
Tobias, thanks you for your kind support in this discussion.
Could you please reference me to some ToF sensors manufactureres you might know?
@Tobias Henzler, Pedro Batista: while I could be wrong of course, TI's ToF sensors or manufactured by Soft Kinetic. Also check out BlueTechnix and the aforementioned ODOS just to name a couple. (Surprisingly, Panasonic has one too, but I don't know if this is still available.)
Here an incomplete list of ToF companies:
Great list, thank you!
The Scorpion 3D Stinger RPP Camera is an alternative to PtGrey and IDS active stereo cameras - read more http://tordivelblog.com/2014/03/22/news-create-3d-images-in-real-time/
The height resolution for the equal to the 2D pixel resolution - the sample and video resolves the volume 300 x 200 x 150 mm. Due to the new Osela RPP laser the system also handles the textureless scenes.
It is slower than TimeOfFlight but is 10-20 more accurate in Z :) - We can supply the camera with resolution from VGA to 5MPixel and use it ourselves for 3D Robot Vision