I'm doctoral student interested to work on these chips for 3-Dimensional position (linear motion) tracking with good precession( <3mm) , accuracy and fast update-rate(>50Hz) . If it is possible for you to design the chips to meet the above requirements, we will place an order as soon as possible. I would like to find an alternative to WII remote .
I am really sorry about my late answer! Actually your question is complex and need more time to answer it. I personally think that, in general, vision sensors are complex detectors and usually need strong knowledge and in particular strong experience in visual processing in specific applications. I am not particularly expert in processing and often I trust to my intuition to design my sensors and sometimes I realize that this is not the correct way to proceed. On the other hand the potentiality of vision sensors is really high. Just think about human eye can do and how many tasks we can perform only thanks to eyes!
Thanks for the comment recently. I did received the Tam2's chip yesterday. I have seen some info about how to use the 5-pins on the rox1 board such as:Vdd(onboard),Gnd(onboard),ANA(onboard),IO6 & IO7 onboard represents CLKIN AND CLB(RESET) respectively if that's the correct order or vice-versa of the two I/O pins onboard. Just like to know if those are 5 pins used for the tam2 chip?, also does ANAOUTBIAS and COLOOUTBIAS pins available on the rox1 board whereas a 1- 10k is tied to them to Gnd for bias purpose. Thanks again! with lots of appreciation.
PS. Are these lens pinhole type, if possible do you know the focal length?.
Hello Geoffrey. Yes, the reason for the resistor was to equalize the gain, not necessarily the offset. In this way, all the control lines to each eye can be paralleled. I was expecting to have to subtract from a white level reading kept in my permanent MRAM. This handles the offset.
I didn't expect to have to change the gain by the scanning software, but doing that handles 'noise' beautifully, especially vertical ghosts... It disappears! I keep a 'black' level in permanent MRAM for each pixel as well as the white level reference. The stored 'black' level is calculated as a fraction to speed things up. The black reference scan uses a darkish target under the same light level as the white reference scan. (Too dark and you make the range too large.)
The fraction is calculated to make the black reference produce a black level 32 pixel counts below the white level. The 8 bit multiplication of this 8 bit fraction with the sampled pixel creates 16 bits PRODH/PRODL in a single cpu cycle and the PRODH byte is the gain-corrected result of a normal scan. Simple and fast.
After that, each picture is brightness adjusted to handle differing lighting conditions. It reduces the 32 level pixel readings to 16 levels, (4 bits) which matches my LCD resolution. I can turn my very-bright lighting LEDs off or leave them on in my living room with only a small difference on the LCD screen.
Amazingly, the fluorescent and LED AC light patterns are not really noticeable... I see a moire fringe every now an then. I think your photsensors do integration just with high impedance RC filtering. I will have to check that sometime.
But I still use the LED strip, 350ma, behind the eyes because they make great reference scans.
So, I have adjusted for offset, gain and brightness .... the resistor allows the signals to be in the right range..important only because I am paralleling the control signals to each eye. I am extremely happy with the results... wish I got a faster cpu, but I have simulated finding the eyes, nose and mouth on my laptop using vertical darkness gradients and this will mean only scanning 1/5 or less of the pixels at high speed, once their location has been identified.
The gradients take pixels 2 above and 2 below the pixel in question and if the real pixel is darker, I multiply the differences by a constant number. The value of the pixel is not used, just the difference between it and the others. If the pixel is good, it is kept, if not, it is set to full white. I then get an average of the non-white pixel values and show as full black only the ones below (darker than) the average. The result is really fun. The constant number can change, the actual value of the pixel or a portion of it can be included and the results are still interesting. Gradient therapy. I will also try varying the 'constant' value in feed forward format to keep the displayed pixels just enough to used for decisions.
I might also change the range of pixels surrounding the pixel in question. I tried this on my laptop and it didn't matter much. I'm looking for edges of horizontal things and the vertical gradient method does this well.
I also tried shifting the whole image left and right to create minumum difference between two eye images. WOW. fun. It works just as well for the gradient screen as the original full screen. Processing this is slow... so gradients would be better.
No comments yet!
You need to be a member of Embedded Eye to add comments!