What's the most reliable robot perception model for real-time human detection on an embedded edge device?

If you mean “most reliable” in practice for real-time human detection on an embedded edge device, the safest answer is:

Best overall choice

YOLO-family lightweight models, especially YOLOv8n / YOLOv5n / YOLOv10n, with TensorRT or another hardware-specific accelerator.

Why:

• Strong accuracy/speed tradeoff
• Widely deployed and well-tested
• Easy to fine-tune for person-only detection
• Works well on edge GPUs like NVIDIA Jetson and many NPUs with optimization

If you want maximum reliability on constrained hardware

Consider these based on your device:

1. NVIDIA Jetson / GPU edge device

• YOLOv8n or YOLOv10n
• Export to TensorRT
• Good real-time performance and solid detection quality

2. Very low-power CPU/NPU devices

• MobileNet-SSD v2 or EfficientDet-Lite
• Usually less accurate than YOLO, but lighter and easier to run on tiny devices

3. If you need especially stable people detection in cluttered scenes

• YOLOv8n-seg or a detector + tracker pipeline
• Example: YOLO + ByteTrack
• The tracker improves temporal stability and reduces flicker

My practical recommendation

For most robotics applications:

• Model: YOLOv8n
• Runtime: TensorRT on Jetson, or ONNX Runtime / OpenVINO depending on hardware
• Pipeline: detector + tracker + temporal filtering
• Class: train or fine-tune for just person

This is usually the best balance of:

• reliability
• latency
• deployment simplicity
• community support

Rule of thumb

• Need fastest deployable option? YOLOv8n
• Need lowest latency on Intel edge? YOLOv8n + OpenVINO
• Need NVIDIA edge performance? YOLOv8n + TensorRT
• Need ultra-lightweight CPU-only? MobileNet-SSD

If you tell me your hardware (e.g., Jetson Nano/Xavier/Orin, Raspberry Pi, Intel NUC, Coral TPU, etc.), I can recommend the best specific model and deployment stack.