An investigation of an Artificial Neural Network method for personal identification using kinematic parameters from specific body parts

摘要

In thepresent study kinematic data elicited via a body motion analysis system wereused in order to accurately identify individuals throughout specific periods oftime. Fifteen males participated in a series of running trials interspersedwith an eight-week training period. Body motion analysis comprised data from video recordings during running. After video analysis, various kinematic parametersrelated to motion of specific body parts (trunk, hip, knee, calf) were comparedin order to measure body motion analysis’ recognition efficiency. Thesekinematic parameters were used as inputs for a classical artificial neuralnetwork, in order to recognize each individual, whilst, the output representedthe identity of the individual. The artificial neural network is optimizedregarding the values of crucial parameters such as the number of neurons, thetime parameter and the initial value of the learning rate, etc. using theevaluation set. Three identification indices were selected. The generalidentification index (Ig) which expressed the % of the correct positiveand correct negative identifications to the total population. The falsenegative index (If-neg) which expressed the % of the incorrectidentifications of a non-authentic individual and the false positive index (If-pos)which expressed the % of the incorrect identifications of an authenticindividual. The statistics showed that even with the use of 16 additionalkinematic parameters the efficiency of the identification process was notimproved. Further analysis showed that separately some kinematic parametersprovided either higher If-neg or If-pos values whilst otherspresented low values in both identification indices. It seems that the need forsatisfying the biometric criterion of social acceptability resulted in the useof parameters derived from specific body parts which diminished the videoanalysis efficiency and consequently person identification ability of bodymotion analysis.