Towards end-to-end gesture recognition with recurrent neural networks

With the development of smart devices, gesture recognition is used in more and more fields. The current gesture recognition devices on the market are inconvenient and expensive. Human motion analysis and recognition based on attitude sensor is a new field. The algorithm based on the recurrent neural network takes into account the timing information of the actions and can better resolve the uncertainty of the human motion in time, but as the training sample increases, the efficiency becomes lower. This paper proposes an action recognition method based on Connectionist temporal classification for sequence learning. This method realizes end-to-end recognition of gestures.