Justin Schoeman wrote:
Gerd Knorr wrote:But my primary question was how to address the capturing of video from two sources. (I'd prefer if it was possibile to use one single thread because (IMHO) it ensures better sync.With v4l2 API you can use select() to do exactly that. GerdWith the v4l2 API, you get timestamps anyway, so you can decide after the fact, which two images are closest in time to each other... No real need for synchronous capture then...
...snip...Since you are "tracking pointers" I would guess that you are translating the information into coordinates for calculation purposes. By time stamping the coordinates from each view point you should be able to interpolate the points from each view. I would pobable use at least a quadratic interpolation using three known points for each view, then generate interpolated data from each veiw at a specific or variable clock speed. Using {x,y,miliseconds} three dimentional points for each pointer could increase accuracy. Storing two sets of three dimentional data sets rather than guestimating the time is more acurate. You may also want to build correction algorithms into your translation, because unless you are using a highspeed shutter the top of each field is "older" than the bottom of the field and the alternate field contains slightly newer information than the first but the data is interlaced below the first fields lines.
One of the most important things to know in order to provide higher accuracy is the technical details of the camera. For example, is the data for each frame captured stop action, or is the data changing while the frame is processed {this is most likely.} What is the frequency drift rate of the camera and can the camera take an external clock. There are other qualities of cameras that are less impotrant, like bluming, bleeding, bluring, pixel variation and the linear acuracy of the optics. These can all be counteracted by doing calibration tests and building a mathematical representaion of the flaws. For instace a spherical aberation in the lens can be calculated and processed with the quadratic representaion of the error. I don't want to give all the solutions, because after all it is your project and your credit not mine. PS I was a math/physics major, but now I just work as a Senior Network Administrator for an Internet Service Provider. Maybe one day I will go back to school and get a masters or doctorate so that I might have a chance getting a job in my prefered field of theoretical physics.
Best advice I can give is to figure out what you want to achive from the most basic aspects of the physical components, calculate representaions of the important aspects {some errors are diminished by other factors.} Then build corrective functions and process the data to achive corrected results. I would suggest it is important to store the raw data, just incase you discover errors in your corrective routines. Reprocessing raw data is more acurate than processing processed data.
As far as the threading of the capture of data, multi threading has advantages if you are maintaining time stamps for each sample. Since the processing of data to gather three dimesioal data to be combined into a four dimentional data stream requires processing, you might just as well build in corrective routines that can provide the required acuracy the project demands. I doubt you want 3D jitter creeping in, that could represent incorrect inertial stability and inaccuratly complex motion analysis. But not knowing your project you may require far less stringent accuacy.
Guy Fraser
