|
|
--- |
|
|
license: apache-2.0 |
|
|
--- |
|
|
Federated Pose HAR is a pose-based human activity recognition dataset designed for evaluating federated learning approaches in industrial and edge-computing scenarios. It contains eight distinct upper-body gestures, represented as sequences of 13 skeletal joint keypoints extracted using a modified FastPose pipeline from RGB-D video recordings. |
|
|
|
|
|
The dataset is organized into: |
|
|
|
|
|
Centralized splits: centralized_x_[train|val|test].txt and centralized_y_[train|val|test].txt, representing aggregated data from all clients for traditional machine learning baselines. |
|
|
|
|
|
Client-wise splits: client_1 through client_5, each containing x_[train|val|test].txt and y_[train|val|test].txt, showing non-IID distributions for federated learning experiments. |
|
|
|
|
|
This structure enables direct, apples-to-apples comparisons between centralized training, local per-client training, standard FedAvg, and ensemble federated setups. |
|
|
|
|
|
Each file in the dataset stores: |
|
|
|
|
|
X files: Normalized skeletal coordinates over time for each gesture sample. |
|
|
|
|
|
Y files: Corresponding gesture class labels (1–8). |
|
|
|
|
|
Data Collection & Preprocessing |
|
|
|
|
|
Gesture videos were recorded in controlled indoor conditions using an RGB-D camera. The skeletal keypoints were extracted with FastPose, modified to output 13 key upper-body joints relevant for industrial gesture control tasks. The extracted 2D keypoint coordinates were normalized relative to torso position and scaled to account for subject height variations. |
|
|
|
|
|
For federated learning experiments, the dataset was partitioned into five non-IID clients, with each client containing samples from a specific subset of gesture instances to mimic realistic heterogeneity. Centralized splits combine all samples across clients for baseline comparisons. |
|
|
|
|
|
This dataset is particularly suited for human activity recognition, pose estimation-based sequence classification, and time-series classification tasks in privacy-sensitive environments. It supports experimentation in multi-client, non-IID scenarios, enabling research into robust FL strategies, personalization techniques, and aggregation methods. |
