Research On Gesture Recognition With Deep Learning Of Inertial Sensor Data

Gesture is an intuitive,natural and convenient way of human-computer interaction.As the technology matures,it is used in more and more scenes,and gesture recognition technology has also become a research hotspot.There are two main types of data collection tools used for gesture recognition: machine vision devices represented by cameras and inertial sensors represented by accelerometers and gyroscopes.Compared with machine vision equipment,inertial sensors have inherent advantages such as low cost,low power consumption,ease of portability,and immunity to external factors such as light.With the continuous development and popularization of micro electromechanical system technology,inertial sensors have been widely used in smart phones and wearable devices.As deep learning has demonstrated excellent performance in a large number of tasks,more and more studies have built gesture recognition systems based on deep learning methods.There are two main research points in this thesis.Among them,research point(1)conducts specific research on the problem of huge parameter quantity of gated neural network,research point(2)conducts specific research on the problems that neural networks are easy to overfit on small data sets and the feature extraction and measurement methods of traditional template matching methods are unreasonable.(1)A single-gated recurrent neural network algorithm is proposed.This algorithm simplifies the double-gated gating mechanism of the gating recurrent neural network into a control mechanism.The memory mechanism is built with a single forget gate as the core,and the contribution of history information and current information to network is comprehensively modulated by forget gate.Experiments and analysis show that the single-gated recurrent neural network reduces the amount of parameters by 42% compared with the gated recurrent neural network,but the reliability and recognition effectiveness are equivalent.(2)An end-to-end gesture recognition algorithm based on Siamese network and hybrid neural network is proposed.The algorithm automatically extracts the spatio-temporal information contained in the inertial sensor signals via a combination of convolutional neural network and recurrent neural network.At the end of the traditional Siamese network branch,two fully connected layers are cascaded as classifiers,the way of feature extraction and similarity measurement can be learned collaboratively during training.Experiments and analysis show that the algorithm has good learning and generalization capabilities.The extracted features not only have strong nonlinear expression ability and separability,but also have certain anti-interference ability,which can handle the variability of gestures.