Chaoticity of gait motion capture data determined by transfer learning in human sex recognition challenge

The application of chaos theory in motion data analysis is the subject of the paper. The main novelty is related to the proposal of a feature extraction technique based on transfer learning and focused on the properties that quantify the level of chaos of non-linear dynamical models. The introduced method is successfully examined in the gait-based human sex recognition challenge. There are two key stages of the proposed approach. In the first one Long Short-Term Memory (LSTM) networks are trained to recognize chaotic and non-chaotic systems based on their univariate observation sequences. They are built using state variables of well-known non-linear dynamical models. Afterward, the network extracts interpretable features of every pose parameter of the human skeleton model. Thus, multivariate motion sequences are decomposed into univariate ones and then processed separately by the LSTM network. As a result, motion descriptors are obtained. Their features are assessed for significant differences between the populations of females and males. Moreover, the performance of the supervised classification is evaluated. In numerical experiments, highly precise motion capture measurements, registered by the gold-standard Vicon system, are used. The obtained results confirm that the proposed approach provides robust and discriminative motion features. They allow for efficient distinction of gait performed by females and males with more than 92% accuracy. Moreover, as statistically significant differences are found and the areas under the receiver operating curve are computed for feature values, it is feasible to assess joints whose movements differ most strongly.