Optimisation of shape kernel and threshold in image-processing motion analysers

The aim of the work is to optimise the image processing of a motion analyse r, This is to improve accuracy, which is crucial for neurophysiological and rehabilitation applications. A new motion analyser, ELITE-S2, for installa tion on the International Space Station is described, with the focus on ima ge processing. Important improvements are expected in the hardware of ELITE -S2 compared with ELITE and previous versions (ELITE-S and Kinelite). The c ore algorithm for marker recognition was based on the current ELITE version , using the cross-correlation technique. This technique was based on the ma tching of the expected marker shape, the so-called kernel, with image featu res. Optimisation of the kernel parameters was achieved using a genetic alg orithm, taking into account noise rejection and accuracy. Optimisation was achieved by performing tests on six highly precise grids (with marker diame ters ranging from 1.5 to 4 mm), representing all allowed marker image sizes , and on a noise image. The results of comparing the optimised kernels and the current ELITE version showed a great improvement in marker recognition accuracy, while noise rejection characteristics were preserved. An average increase in marker co-ordinate accuracy of +22% was achieved, corresponding to a mean accuracy of 0.11 pixel in comparison with 0.14 pixel, measured o ver all grids. An improvement of +37%, corresponding to an improvement from 0.22 pixel to 0.14 pixel, was observed over the grid with the biggest mark ers.