Development of Equine Gait Recognition Algorithm

Abstract: Horseback riding is a sport enjoyed by people around the world. Many riders are interested in knowing exactly how much they have exercised their horse and how much time that have been spent in different gaits. The goal of this master's thesis was to develop an equine gait recognition algorithm. Triaxial accelerometer and gyroscope signals were collected during different riding sessions by using smartphones. Features, used in previous activity recognition works, were implemented and calculated for all sensor signals. Different methods to select important features were used and the feature sets were then evaluated. In the work four classifiers were implemented and evaluated. The work resulted in an equine gait recognition algorithm based on signals collected at the saddle-girth. The developed algorithm used a window length of 128 samples and windows with 50 % overlap. A feature set was chosen by the use of sequential forward feature selection. Five features were included in the final algorithm and two classifiers using two respectively three of the features. The first classifier separated stand from the gaits by using the features root mean square for the magnitude of the gyroscope signal and energy of the x-axis accelerometer signal. The second classifier classified gaits as either walk, trot or canter using the wavelet based feature energy distribution ratio of the z-axis accelerometer signal, dominant frequency of z-axis of the gyroscope signal and skewness of the accelerometer z-axis. The classifiers used in both classification steps were KNN with K = 3. The algorithm performed well on a collected test set including two riding sessions. It should be noted that the same phone was used to collect both training and testing data. The performance of the developed algorithm was benchmarked against the smartphone application Equilab. The performance of both algorithms was similar. The developed equine gait recognition algorithm had a 94.1 % and 97.4 % accuracy on the two different test sessions. Further development of the algorithm will be needed to include other terrains and a larger variety of horses and riders.