Accelerating embedded image processing for real time: a case study

A Practical Multirobot Localization System

“…The average processing time per image for each dataset was measured and calculated by (20). The results summarized in Table 5 indicate the correctness of the model described in Section 5.3.…”

“…The pattern symmetry is exploited to perform the position and orientation estimation separately, which allows to base the pattern localization on a two-dimensional convolution. Although this convolution-based approach has proven to be reliable enough to achieve 24/7 operation, its computational complexity still remains high, which lead to its implementation on alternative platforms such as FPGA [20].…”

“…Moreover, the smaller the pattern image dimension, the faster the processing rate. Of course, equation (20) gives only a coarse estimate, but it might give the user a basic idea of the system processing speed. Equation (20) has been experimentally verified and the results are presented in Section 6.3.…”

“…Thus, the hypothesis is to test if the algorithm processing speed estimation (20) conforms to the proposed assumptions. Moreover, in this experiment, we also verify if the algorithm complexity depends only on the pattern size rather than on the image resolution.…”

“…Thus, one should be able to establish what kind of computational hardware is needed for a particular setup. This can be roughly estimated using the time to process one image by means of (20). Three real world datasets and five different platforms, including two credit-card sized computers, were used to verify the model in a realistic setup.…”