3D Dual Path Networks and Multi-scale Feature Fusion for Human Motion Recognition
Authors:
- 1 December 2019
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 1, 698-704
Abstract
In this paper, a novel 3D dual path networks (3D-DPN) with 3D convolutional block attention module (3D-CBAM) and multi-scale feature fusion (MFF) was established. The architecture converted all 2D operations in the DPN structure into 3D and combined with 3D-CBAM to make the network have attention mechanisms for channel and spatiotemporal information. In order to keep the network sensitive to spatiotemporal input on different scales, we integrate the output of each 3D-DPN+CBAM block at the top of the network. This multi-scale feature fusion operation also allows the network to expand flexibly to accommodate different input frame lengths without significantly increasing the number of parameters and computation. We utilize this structure for human motion recognition based on the depth-sensing video. The experimental results indicate that our method higher-quality than the state-of-the-art model on multiple datasets.Keywords
This publication has 19 references indexed in Scilit:
- Combining CNN streams of RGB-D and skeletal data for human activity recognitionPattern Recognition Letters, 2018
- Action recognition based on joint trajectory maps with convolutional neural networksKnowledge-Based Systems, 2018
- Attention-Based Multiview Re-Observation Fusion Network for Skeletal Action RecognitionIEEE Transactions on Multimedia, 2018
- Densely Connected Convolutional Networks2017
- Aggregated Residual Transformations for Deep Neural Networks2017
- Deep Residual Learning for Image Recognition2016
- Dynamic Image Networks for Action Recognition2016
- Hierarchical transfer learning for online recognition of compound actionsComputer Vision and Image Understanding, 2016
- UTD-MHAD: A multimodal dataset for human action recognition utilizing a depth camera and a wearable inertial sensor2015
- Two-person interaction detection using body-pose features and multiple instance learning2012