Deep Convolutional and LSTM Neural Network Architectures on Leap Motion Hand Tracking Data Sequences

This paper focuses on the hand gesture recognition problem, in which input is a multidimensional time series signal acquired from a Leap Motion Sensor and output is a predefined set of gestures. In the present work, we propose the adoption of Convolutional Neural Networks (CNNs), either in combination with a Long Short-Term Memory (LSTM) neural network (i.e. CNN-LSTM), or standalone in a deep architecture (i.e. dCNN) to automate feature learning and classification from the raw input data. The learned features are considered as the higher level abstract representation of low level raw time series signals and are employed in a unified supervised learning and classification model. The proposed CNN-LSTM and deep CNN models demonstrate recognition rates of 94% on the Leap Motion Hand Gestures for Interaction with 3D Virtual Music Instruments dataset, which outperforms previously proposed models of handcrafted and automated learned features on LSTM networks.