Dual-Handed Dynamic Gesture Recognition using Inertial Sensors

Dual-handed gesture recognition is investigated using inertial measurement units (IMUs). A database of 10 types of dual-handed dynamic gestures is collected by five people with 250 samples each. The training set consists of 2,000 inertial data, and the test set consists of 500. A three-layer convolutional neural network (CNN) is used as the main training model to recognize gesture movements. The collected row wave data is directly sent to CNN for deep learning, and the convolutional layer extracts the IMU features of different gestures. Combined with the activation function (ReLU) and efficient gradient descent and backpropagation, the problems of gradient issues during training are avoided. The intermediated output data is compressed by the pooling layer (Max pooling), and the collected data from both hands is merged using the concatenate method. Finally, the fully connected layer is used for feature classification. The final success rate of 97.5% is achieved in the designed system.