Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Atchitectures of deep neural network for gait recognition and human pose estimation with pre-trained models
| Date | Stars |
|---|---|
| 2026-07-31 | 369 |
| 2026-08-05 | 369 |
| 2026-08-06 | 369 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# DNN for gait recognition in TensorFlow
In this project you can find implementation of deep neural network for **people identification from video** by the characteristic of their **gait**. The processing is very robust against various covariate factors such as *clothing*, *carrying conditions*, *shoe types* and so on. Feel free to use this network in your project or extend it in some way.
## Requirements
The code was written in `Python 3.5`, but it is probably also compatible with other versions.
### Python packages
- `TensorFlow 1.3` - [how to install](https://www.tensorflow.org/install/) (however, some parts of code may be compatible with lower version, report an issue if you would have any problems.)
- `numpy`, `scipy`, `PIL`
## Basic information about architecture
The network takes *raw RGB video frames* of a pedestrian as an input and produces one-dimensional vector - **gait descriptor** that exposes as an identification vector. The identification vectors from gaits of each two different people should be **linearly separable**. Whole network consists of two sub-networks connected in cascade - `HumanPoseNN` and `GaitNN`.
**Spatial features** from the video frames are extracted according to the descriptors that involve **pose of the pedestrian**. These descriptors are generated from the first sub-network - `HumanPoseNN` defined in `human_pose_nn` module. `HumanPoseNN` can be also used as a standalone network for regular **2D pose estimation problem** from still images (for more info see [this section](#pose-estimation)).
Responsibility of the second sub-network - `GaitNN` defined in `gait_nn` module is the further processing of the generated spatial features into one-dimensional **pose descriptors** with the use of a residual convolutional network. **Temporal features** are then extracted across these *pose descriptors* with the use of the multilayer recurrent cells - **LSTM** or **GRU**. All temporal features are finally aggregated with **Average temporal pooling** into one-dimensional **identification vector** with good discriminatory properties. As already mentioned in the text above, the human identification vectors are linearly separable with each other and can therefore be classified with e.g. **linear SVM**.
## Gait recognition
The dummy code bellow shows how to generate the identification vector form the input data `video_frames`. For the best results, all frames should include the **whole** person visible from the **profile view**. The person should be located approximately in the center of each frame.
```python
# Initialize computational graphs of both sub-networks
net_pose = HumanPoseIRNetwork()
net_gait = GaitNetwork(recurrent_unit = 'GRU', rnn_layers = 2)
# Load pre-trained models
net_pose.restore('path/to/pose_checkpoint.ckpt')
net_gait.restore('path/to/gait_checkpoint.ckpt')
# Create features from input frames in shape (TIME, HEIGHT, WIDTH, CHANNELS)
spatial_features = net_pose.feed_forward_features(video_frames)
# Process spatial features and generate identification vector
identification_vector = net_gait.feed_forward(spatial_features)
```
## Pose estimation
The first sub-network `HumanPoseNN` can be also used as a standalone network for 2D **pose estimation problem**. This can be done in such a way:
```python
net_pose = HumanPoseIRNetwork()
# Restore pre-trained model
net_pose.restore('path/to/pose_checkpoint.ckpt')
# input_images should contain RGB images (299 x 299) to be processed.
# The images in batch should be stacked along the first dimension, so the shape of input_images
# has to be (BATCH, 299, 299, 3)
coords_y, coords_x, probabilities = net_pose.joint_positions(input_images)
```
where `coords_y`, `coords_x` and `probabilities` stores estimated joint coordinates in **Y axis**, **X axis** and **probability** of each estimate, respectively. All these tensors have shape `(BATCH, 16)`, where the second dimension is the body joint. The Excerpt of 9,262 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:009976bed26873fd, llm:description: 'Atchitectures of deep neural network for gait recognition and human pose estimation with pre-trained models' (repo description)
matched fp:009976bed26873fd, llm:description: 'Atchitectures of deep neural network for gait recognition and human pose estimation with pre-trained models' (repo description)
matched fp:009976bed26873fd, llm:description: 'Atchitectures of deep neural network for gait recognition and human pose estimation with pre-trained models' (repo description)