Skeleton-based Human Activity Classification in Sparse Image Sequences

Abstract

Research results on human activity classification in video are described, based on initial human skeleton estimation in video frames. Both single person actions and two-person interactions are considered. The initial skeleton data is estimated in selected video frames by OpenPose, HRNet or other dedicated library. Important contributions of presented work are computational steps of skeleton tracking and -refinement, and relational feature extraction from pairs of skeleton joints. It is shown, that this feature engineering significantly increases the classification accuracy. Regarding the final neural network encoder-classifier, two different architectures are designed and tested. The first solution is a lightweight MLP network, implementing the idea of a "mixture of pose experts". Several pose classifiers (experts) are trained independently on different time periods (snapshots) of single-person visual actions (or 2-person interactions), while the final classification is a time-related pooling of weighted expert classifications. All pose experts use the same deep encoding network. The second (middle weight) solution is based on a LSTM network.
Both solutions are trained and tested on the action set of the well-known NTU RGB+D dataset, although only 2D data are used.
Our results show comparable performance with some of the best reported STM- and CNN-based classifiers for this dataset. We conclude that by reducing the noise of skeleton data, highly successful lightweight- and midweight-approaches to visual activity recognition in image sequences can be achieved.