Anomaly Detection of Motion Capture Data Based on the Autoencoder Approach

Anomalies of gait sequences are detected on the basis of an autoencoder strategy in which input data are reconstructed from their embeddings. The denoising dense low-dimensional and sparse high-dimensional autoencoders are applied for segments of time series representing 3D rotations of the skeletal body parts. The outliers - misreconstructed time segments - are determined and classified as abnormal gait fragments. In the validation stage, motion capture data registered in the virtual reality of the Human Dynamics and Multimodal Interaction Laboratory of the Polish-Japanese Academy of Information Technology equipped with Motek CAREN Extended hardware and software are used. The scenarios with audio and visual stimuli are prepared to enforce anomalies during a walk. The acquired data are labeled by a human, which results in the visible and invisible anomalies extracted. The neural network representing the autoencoder is trained using anomaly-free data and validated by the complete ones. AP (Average Precision) and ROC-AUC (Receiver Operating Characteristic – Area Under Curve) measures are calculated to assess detection performance. The influences of the number of neurons of the hidden layer, the length of the analyzed time segments and the variance of injected Gaussian noise are investigated. The obtained results, with AP = 0.46 and ROC-AUC = 0.71, are promising.