[VFS 2019] Human Activity Recognition Approaches
Download as PPTX, PDF0 likes236 views
This document provides an overview of approaches for human activity recognition (HAR). It discusses vision-based approaches including using hand-crafted motion features, depth information methods, and deep learning techniques. Deep learning methods covered include two-stream inflated 3D networks, deep bidirectional LSTM with CNN features, and skeleton-based spatial temporal graph convolutional networks. The document also discusses potential future directions for HAR, which may involve using 3D features, unsupervised learning, skeleton-based methods for privacy, and optimization/quantization. HAR applications in the future could leverage the power of modern AI and embedded abilities while meeting requirements like real-time performance, privacy, and ease of deployment.
[VFS 2019] Human Activity Recognition Approaches
- 1. —Lao Tzu “The journey of a thousand miles begins with one step.”
- 2. Human Activity Recognition Approaches Nguyễn Việt Thái / CTO of Asilla Vietnam Vietnam Frontier Summit 2019
- 3. Some of the products of ... Our clients include...
- 4. Introduction Approaches for Human Activity Recognition HAR in the future Some applications of human activity recognition Vision Based approaches We are doing, Approaches in future 01 02 03
- 7. Some applications of HAR
- 8. Some dataset Name Number of Data Categories UCF101 13,320 101 HMDB51 6,849 51 MSRAction3D 20 ActivityNet 27,811 203 AVA 430 80 NTU RGB+D 56,880 60 NTU RGB+D 120 11,448 120
- 11. Device selection Surveillance camera Data collection, RGB, RGB-D Preprocessing, Feature selection Machine Learning / Deep learning The general process of HAR
- 12. Vision Based Using Hand- crafted motion features Depth information based methods Deep learning based methods Approaches
- 13. Using Hand-Crafted Motion Features ● Key-framing ● Spatial-temporal features ● Space-time interest points ● Histograms of optical flow.
- 14. Depth Information Based Methods ● RGBD ● The interest of applying depth data captured from depth cameras for the action recognition
- 15. Deep Learning Based Methods ● CNN ● Long-Short Term Memory ● Two-Stream Architecture ● Skeleton Based
- 16. Deep Learning Based Methods Deep Bi- Directional LSTM With CNN Features Long-term recurrent convolutional networks Temporal Segment Network (TSN) Two-Stream Inflated 3D(I3D) 15 16 Long-term temporal convolutions 1817 Skeleton Based
- 18. DB-LSTM With CNN Features
- 19. Skeleton Based Spatial Temporal Graph Convolutional Networks
- 20. Pose Estimation ● Top-Down approaches ○ Alphapose: MAP: 78.6% on Coco, FPS: ~20 ○ Mask R-CNN: MAP: 75.6% on Coco, FPS: ~3 ● Bottom-Up approaches ○ Deepercut: MAP: 70% on MPII Multi-Person, FPS: ~15 (myself) ○ Openpose: MAP: 68.2% on Coco, FPS ~ 21.7 with body+ ● Other ○ Posefix: Increate 3~8% for each approaches
- 21. Benchmark
- 22. HAR in the future 03
- 23. Depth Camera The power of modern AI Embedded Ability
- 24. Requirements ● Ability to deploy on compact devices. ○ Camera, hardware device compact. ● Ability to run realtime. ● Ability to secure user information. ● Easy to deploy in different environments.
- 25. Potential ● Use 3D Features ● Unsupervised learning. ● Skeleton Based for safe in privacy ● Optimization & quantization
- 26. Does anyone have any questions? (+84) 0988144281 thai@asilla.net THANKS!
- 27. Instructions for use 1. Different Approaches for Human Activity Recognition– A Survey Zawar Hussain, Michael Sheng, Senior Member, IEEE, Wei Emma Zhang, Member, IEEE 2. Going Deeper into Action Recognition: A Survey. Samitha Herath, Mehrtash Harandi, Fatih Porikli 3. Video Activity Recognition: State-of-the-Art. Itsaso Rodríguez-Moreno, José María Martínez-Otzeta, Basilio Sierra, Igor Rodriguez andEkaitz Jauregi. 4. This document is the official template of presentation at Vietnam Frontier Summit 2019 (VFS) and will only be used for purposes relating to the event. 5. Temporal Segment Networks. Limin Wang1 , Yuanjun Xiong2 , Zhe Wang3 , Yu Qiao3 , Dahua Lin2 , Xiaoou Tang2 , and Luc Van Gool1 6. Long-term Temporal Convolutions for Action Recognition. Gul Varol, Ivan Laptev, and Cordelia Schmid, ¨ Fellow, IEEE