SensiFoot V8 Edge Models
This repository contains the official model weights for the SensiFoot SDK architecture, an industrial, edge-optimized pipeline for real-time human lower-limb gesture recognition.
These models are provided to support the code availability and reproducibility of our upcoming academic publications.
🔗 Code Repository
This repository contains the files which have been placed at that github repo! To run these models and view the full implementation, please use the official SDK and inference engine: GitHub: Sensifai-BV/sensifoot-sdk
Model Architecture Details
The SensiFoot V8 pipeline eschews heavy CNN-BiLSTM hybrids in favor of a highly optimized, single LSTM network with a temporal attention layer. This design allows for near-zero latency on edge hardware.
- Input Features: The model ingests a 36-dimensional feature set (Phase 5 configuration), encompassing extracted positional velocities and joint angles. This specific feature engineering ensures resilience against monocular depth jitter.
- Gating Mechanism: Live inference utilizes Adaptive Ordinal Distance (A-OD) gating to establish a dynamic mathematical noise floor ($\mu + 3\sigma$), autonomously separating idle states from active gesture execution.
- Camera Views: The feature extraction workflow is optimized for front-facing camera perspectives.
Usage
Do not load these models directly using standard PyTorch/ONNX boilerplate unless you are modifying the architecture. Please use the SensiFoot Engine provided in the GitHub repository to ensure the EMA-smoothed feature extraction and A-OD gates function correctly.
Citation
(Citation details will be updated upon publication in Engineering Application of Artificial Intelligence / Knowledge-Based Systems).