AMD Kria K26 SOM runs privacy-preserving fall detection at 4.5 FPS
On-device pose estimation detects falls without any cloud dependency or video feeds.
A team from San Diego State University developed a compact, battery-operated fall prediction and detection system running on the AMD Kria K26 System-on-Module (SOM). Using an Intel RealSense D455 range-sensing camera (640x480 depth at 60 FPS), the three-stage pipeline first runs a quantized YOLOX model to detect human bounding boxes from RGB frames — then immediately discards the RGB data to preserve privacy. The A2J (Anchor-to-Joint) model estimates 15 keypoints from depth frames, and a convolutional neural network classifies fall activity from selected joint coordinates (x, y, z).
The system was tested with a single-core DPU serial pipeline (2.5 FPS) and a dual-core DPU multi-threaded version (4.5 FPS), achieving a classification accuracy of 75.85% on the UR Fall Detection and custom SDSU PSG datasets. The entire inference pipeline runs locally on the edge device, eliminating any need for cloud connectivity. This makes the system particularly well-suited for assisted living facilities and home monitoring where both latency and privacy are critical. Future work aims to improve model accuracy and increase throughput beyond 4.5 FPS.
- Uses quantized YOLOX for bounding box detection, A2J for 15-joint pose estimation, and a CNN for fall classification — all on AMD Kria K26 SOM
- Achieves 4.5 FPS throughput with dual-core DPU multi-threaded pipeline and 75.85% classification accuracy
- RGB frames are discarded after bounding box detection to ensure privacy; only depth data is used for pose estimation and fall detection
Why It Matters
Edge-based fall detection with full privacy and no cloud dependency is now feasible at low power and low cost.