Combined Spatial Features Recognition Method for Similar Video Action Automatic Classification

Due to the differences between general actions are obvious, these actions are easily recognized by traditional methods which are input with RGB data. But for similar actions, the differences between them are subtle. Especially when the background moves violently or the light is unstable, RGB data is not robust, and recognizing similar actions remains difficult. To improve the recognition accuracy of similar actions, the skeleton-based features and the difference between similar actions are focused on. In this paper, we proposed the FAU-Bi-LSTM algorithm, which could accurately recognize four types of similar activities (walking, running, riding bike, climbing stairs) based on the relative angle feature and relative distance feature and feature attribute unit, and these features could preserve the spatial difference information. The experiment results show that the FAU-Bi-LSTM algorithm could recognize these four types of similar activities by their SDC(Spatial-Difference-Contained) features and has a better performance in similar activities recognition.