Análise do padrão cinemático da marcha em equinos por meio de redes neurais artificiais

Abstract

A total of 39 adult horses (29 male and 10 female), gait champions and reserved champions of Campolina, Mangalarga, Mangalarga Marchador and Pampa breeds at national expositions, were analyzed by the experiment described in chapter II. For characterization of the gait pattern of the different breeds, 8 different joint angles, obtained by video film digitalization were processed by artificial neural nets (ANN) with multi-objective algorithm (MOBJ) and back propagation (BP). The video frequency was 200 Hz and the digitalization software was Simi Motion 3D, version 7.2. The movement pattern of each animal was obtained by the analysis of three step cycles. As input information for the ANNs was calculated the variable delta as the difference between the maximum and minimum value of each joint angle. The routines, implemented by Matlab 7.0 software, were executed 100 times in order to enable a mean performance of classification. The mean accuracy for breed identification was 98.4% for MOBJ and 93.0% for BP. For cross validation MOBJ resulted in a better performance (95.1%) than BP (84.1%) concerning the generalization, accuracy statistics, sensibility and specificity, because the MOBJ simultaneously minimize of the error and the norm weights by MOBJ and BP only minimize mean square error. 26 adult horses (26 male and 10 female) of Mangalarga breed, participants of the national exposition, were analyzed by the experiment described in chapter III, to evaluate the applicability of ANN for the prediction of step length. 28 linear measurements and 8 joint angles obtained by video film analysis were used for analysis. The objective was to characterize the gait pattern and to compare the algorithms multi-objective and multi-objective LASSO (MOBJ-LASSO). The movement pattern of each animal was obtained by the analysis of three step cycles. The input variables of the ANN were the variable delta (difference between maximum and minimum joint angles) and the 28 linear measurements. The mean accuracy for step length prediction was 98.3% for MOBJ and 97.2% for MOBJ-LASSO. For cross validation MOBJ (96.0%) and MOBJ-LASSO (96.5%) showed similar performance. The MOBJ-LASSO is a better algorithm than MOBJ, it is able to eliminate inputs and to perform an automatic selection of parameters for the ANN. That way, the ANNs become more effective for the identification of relevant variables. The MOBJ-LASSO routine selected four (three joint angles and one linear measurement) of the 36 variables (28 linear measurements and 8 joint angles) to better define the step length.