CNN-based and DTW features for human activity recognition on depth maps

Abstract In this work, we present a new algorithm for human action recognition on raw depth maps. At the beginning, each class train separate one-against-all convolutional neural network (CNN) to extract class-specific features representing person shape. Each class-specific, multivariate time-series is processed by Siamese multichannel 1D CNN or determine actions. Afterwards, nonzero pixels shape in map calculate statistical features. On of such Dynamic Time Warping (DTW) They are determined basis DTW distances between all training time-series. Finally, feature vector concatenated with vector. For category multiclass classifier, which predicts probability distribution labels. From pool classifiers select number that an ensemble built them achieves best classification accuracy. Action performed soft voting averages distributions calculated largest discriminative power. We demonstrate experimentally MSR-Action3D and UTD-MHAD datasets proposed attains promising results outperforms several state-of-the-art depth-based algorithms.