Activity Recognition

lastmod 2025-02-11created 2020-08-0422 edits

Action recognition, or activity recognition, aims to recognize the actions and goals of one or more agents from a series of observations of the agents' actions and the environmental conditions.

Video Action Recognition

MXNet: TSN, C3D, I3D, I3D_slow, P3D, R3D, R2+1D, Non-local, SlowFast
PyTorch: TSN, I3D, I3D_slow, R2+1D, Non-local, CSN, SlowFast, TPN

Recognizes human actions in video.

Action Recognition in Video: Exploration of different solutions to action recognition in video, using neural networks implemented in PyTorch.; https://github.com/eriklindernoren/Action-Recognition

Temporal Action Detection

Temporal Segment Network (TSN): Splits the whole video into N equal parts and samples a random snippet (clip) within each of the N parts; A model that passes those snippets through two networks, Spatial & Temporal, and then combines the results of all snippets to produce the output; Temporal Segment Network (TSN)

Temporal Pyramid Network for Action Recognition (TPN): https://decisionforce.github.io/TPN/

SlowFast: SlowFast Networks review

Human Activity Recognition

In the field of Human Activity Recognition (HAR), the first stage can be divided into the processes of "recognizing" and "classifying" human actions.

After going through the process of recognizing a human action, it must be labeled appropriately to classify what the action is. For example, "walking", "running", etc.
In the end, in the Activity Classification Approach, each activity label is attached in the final result through a Classification algorithm.

Template-based approach: This is simple but computationally expensive.

Generative model: Given data x and its classes y as input, a Generative model first learns a model of the joint probability P(x|y) of input x and label y, then computes P(y|x) using Bayes rule, and finally the algorithm selects the most likely label y.; Because it models the distribution of data (x) for each class (y), it can generate sample data.; Examples of Generative models include the Gaussian mixture model, Hidden Markov model, and Dynamic Bayesian Network.

Discriminative model: Unlike the Generative model, the Discriminative model learns P(y|x) directly, so it uses this value directly to classify the class. That is, for a new input, it computes the posterior of every class and decides on the class with the highest probability.; Examples of Discriminative models include the Support Vector Machine, Neural networks, and Relevance Vector Machine.

Deep Human Action Recognition: https://github.com/dluvizon/deephar

2D/3D Pose Estimation and Action Recognition using Multitask Deep Learning (Youtube video): https://www.youtube.com/watch?v=MNEZACbFA4Y&t=6s

Skeleton Based Action Recognition

Skeleton Based Action Recognition | Papers With Code

MMSkeleton: A OpenMMLAB toolbox for human pose estimation, skeleton-based action recognition, and action synthesis.; https://github.com/open-mmlab/mmskeleton; Mmskeleton_demo_video.gif

Multi-person Real-time Action Recognition Based-on Human Skeleton: Apply ML to the skeletons from OpenPose; 9 actions; multiple people. (WARNING: I'm sorry that this is only good for course demo, not for real world applications !!! Those ary very difficult !!!); https://github.com/felixchenfy/Realtime-Action-Recognition; Realtime-Action-Recognition-demo-video.gif

Residual Graph Convolutional Network (ResGCN) v1.0: ResGCN - an efficient baseline for skeleton-based human action recognition.; https://github.com/yfsong0709/ResGCNv1

Documentation

Independently Recurrent Neural Network (IndRNN) - Building A Longer and Deeper RNN: https://arxiv.org/abs/1803.04831

Deep Independently Recurrent Neural Network (IndRNN): https://arxiv.org/abs/1910.06251

Favprite site

References

Awesome-action-recognition-README-ef12a5c.md.zip ↩

author Zer0 sitemap