Research On Mems Sensor-Based Gesture Recognition Algorithm

In recent years, due to the rapid development of virtual technology, researching human-computer interaction has gradually become to be a hot spot, and gesture recognition has a very important position in human-computer interaction. Gesture recognition is including vision-based gesture recognition and sensor-based gesture recognition, while the vision-based recognition is very strict with surrounding and hardware equipment. With the rapid development of MEMS sensor technology, sensor-based gesture recognition has become to be the current research focus. This paper will be based on MEMS sensors, study gesture recognition algorithm.MEMS sensors can provide sensor data but cannot provide relevant data about gesture motion directly. Therefore, the general sensor-based gesture recognition algorithm is setting action firstly, and then analyzes their acceleration or angular velocity and other information. The purpose is to find the key information of each gesture by kinematic characteristics of it. In order to make the operator more flexibility to perform gestures, as well as they can define their own gestures, we use a method called spatial localization to convert MEMS sensor data into spatial data, in order to record motion coordinates of the gesture in three-dimensional space.For the classification of spatial data obtained by integrating gesture, we use Restricted Boltzmann Machine to extract the key information data, which is one of deep learning algorithms. Namely reduces the dimensions of the data. And then use the BP neural network to classify. In order to extract more abstract feature information, this paper applies the deep belief network model, which is consisted by multiple Restricted Boltzmann Machine and BP neural network. Training data samples layer by layer using Restricted Boltzmann Machine, to get information of dimensionality reduction, and then use BP algorithm to adjust the entire network and to classify.The results of experiment show that the conversion of spatial data can accurately record the trajectory of the gestures in three-dimensional space, implant MEMS sensor data into spatial data and then based on this classification is entirely feasible. Training the spatial data of gesture by deep learning algorithms can avoid artificial participation, and reach the purpose of the operator can define their own gestures.