Action Recognition Based on Multi-Layer Visual Features

Action recognition is a crucial problem in computer visual. Conventional methods mainly focuses on spatio-temporal features with intensively computational overhead, while we propose an efficient method recognizing action with fusing sequential visual features. The proposed method combines the motion information across frames and apparent information using a new feature coding layer. The motion information are encoded with dynamic pose flow field, which is based on human pose trajectory. We obtain the sequential visual features by end-to-end learning, and classify the features with SVM. Through experiments on public databases such as KTH database, Penn database and HMDB-51 database, we demonstrate the effectiveness of our approach.