研究目的
To propose a real-time 3D human pose forecasting system using deep neural network to predict human’s future pose from normal RGB frames and visualize the prediction result on VR head-mounted display.
研究成果
The system is capable of 3D pose estimation and forecasting in virtual reality, applicable to real-time visualizations, sports estimations, or trainings. Future work includes evaluation of inference time and accuracy tests.
研究不足
The method has not been evaluated for inference time or accuracy of the forecasted poses, which are areas for future work.
1：Experimental Design and Method Selection
The system consists of three main components: 2D pose estimation from RGB frames using a residual neural network, 2D pose forecasting using a recurrent neural network, and 3D recovery from the predicted 2D pose using a residual linear network. A novel lattice optical flow method is proposed for joint movement estimation.
2：Sample Selection and Data Sources
The system uses annotated 3D human pose datasets for training. Input consists of 224 × 224-size cropped RGB images.
3：List of Experimental Equipment and Materials
RGB camera, VR head-mounted display.
4：Experimental Procedures and Operational Workflow
1. 2D pose estimation from RGB frames. 2. 2D pose forecasting using LSTM network with lattice optical flow. 3. 3D pose recovery from predicted 2D poses.
5：Data Analysis Methods
The system uses a customized ResNet50 for 2D pose estimation, LSTM networks for pose forecasting, and a residual linear network for 3D recovery. A novel lattice optical flow method is used to reduce computation.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
