Feature Extraction and Pattern Recognition for Human Motion by a Deep Sparse Autoencoder

Abstract

Human motion data is high-dimensional time-series data, and it usually contains measurement error and noise. Recognizing human motion on the basis of such high-dimensional measurement row data is often difficult and cannot be expected for high generalization performance. To increase generalization performance in a human motion pattern recognition task, we employ a deep sparse auto encoder to extract low-dimensional features, which can efficiently represent the characteristics of each motion, from the high-dimensional human motion data. After extracting low-dimensional features by using the deep sparse auto encoder, we employ random forests to classify low-dimensional features representing human motion. In experiments, we compared using the row data and three types of feature extraction methods - principal component analysis, a shallow sparse auto encoder, and a deep sparse auto encoder - for pattern recognition. The experimental results show that the deep sparse auto encoder outperformed the other methods with the highest average recognition accuracy, 75.1%, and the lowest standard deviation, ±3.30%. The proposed method, application of a deep sparse auto encoder, thus enabled higher recognition accuracy, better generalization and more stability than could be achieved with the other methods.