Learning Spatiotemporal Features Using 3DCNN and Convolutional LSTM for Gesture Recognition

TLDR

Gesture recognition aims to understand ongoing human gestures. This paper proposes a deep architecture to learn spatiotemporal features for gesture recognition. The architecture first extracts 2D spatiotemporal maps with 3DCNN and bidirectional ConvLSTM, then refines them with 2DCNN while preserving spatiotemporal correlations. The resulting 2D maps encode global temporal and local spatial information, enabling effective gesture recognition, and experiments on IsoGD and SKIG datasets demonstrate its superiority.

Abstract

Gesture recognition aims at understanding the ongoing human gestures. In this paper, we present a deep architecture to learn spatiotemporal features for gesture recognition. The deep architecture first learns 2D spatiotemporal feature maps using 3D convolutional neural networks (3DCNN) and bidirectional convolutional long-short-term-memory networks (ConvLSTM). The learnt 2D feature maps can encode the global temporal information and local spatial information simultaneously. Then, 2DCNN is utilized further to learn the higher-level spatiotemporal features from the 2D feature maps for the final gesture recognition. The spatiotemporal correlation information is kept through the whole process of feature learning. This makes the deep architecture an effective spatiotemporal feature learner. Experiments on the ChaLearn LAP large-scale isolated gesture dataset (IsoGD) and the Sheffield Kinect Gesture (SKIG) dataset demonstrate the superiority of the proposed deep architecture.