skills.rest
SharpAI
SharpAIOFFICIAL
Explore all skills by this author →
GitHub

yolo-detection-2026-coral-tpu-macos

Detects objects in camera frames using Google Coral Edge TPU hardware acceleration.

3.0k|478|Updated Mar 5, 2019
Search Tags:#object detection#surveillance#yolo#edge inference#TFLite#Real-time#coral edge tpu
One-click install
npx skills add https://github.com/SharpAI/DeepCamera --skill yolo-detection-2026-coral-tpu-macos
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: yolo-detection-2026-coral-tpu-macos
Source: https://github.com/SharpAI/DeepCamera/tree/main/skills/detection/yolo-detection-2026-coral-tpu-macos
Command: npx skills add https://github.com/SharpAI/DeepCamera --skill yolo-detection-2026-coral-tpu-macos

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires ai-edge-litert, numpy, Pillow, and includes scripts (resource) components.

What problem does it solve?

Running real-time object detection on live camera streams typically requires a GPU or cloud inference, which adds cost, latency, and privacy concerns. This skill offloads detection to a Google Coral Edge TPU USB accelerator, delivering ~4ms inference at 320x320 resolution entirely offline on macOS and Linux.

Core Features & Use Cases

  • Hardware-Accelerated Detection: Runs INT8-quantized TFLite models on the Coral Edge TPU via the ai-edge-litert runtime, detecting 80 COCO classes (person, car, dog, cat, etc.).
  • TPU Health Watchdog: Detects USB hangs with a 10-second invoke timeout and identifies thermal stalls after 30 consecutive empty frames, emitting telemetry events for automatic recovery.
  • CPU Fallback: Gracefully falls back to CPU inference with an SSD MobileNet model when no Edge TPU is detected.
  • Use Case: Monitor a front-door camera on a Mac mini by detecting people and cars in real time at 5-15 FPS without sending any footage to the cloud.

Quick Start

Run the deploy.sh script to install the Edge TPU drivers and Python environment, then ask the agent to start detecting people and cars on your camera stream.

Frequently Asked Questions about yolo-detection-2026-coral-tpu-macos

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I run object detection on a Google Coral Edge TPU?▼

Run the deploy.sh script, which installs libedgetpu drivers, creates a Python virtual environment, and downloads compiled TFLite models. The detect.py script then loads the Edge TPU delegate via ai-edge-litert and processes frames sent over stdin as JSONL events.

What inference speed does Coral Edge TPU achieve for object detection?▼

At 320x320 input resolution, inference takes approximately 4ms and runs fully on the TPU. At 640x640, inference takes around 20ms because some model layers are offloaded to the host CPU.

Does Coral Edge TPU work on Apple Silicon macOS?▼

Yes, Apple Silicon is supported natively using the feranick/libedgetpu community fork, which provides an arm64 dylib without requiring Rosetta 2. Intel Macs use the official Google Edge TPU runtime installer instead.

What happens if no Edge TPU is connected during detection?▼

The skill falls back to CPU inference automatically. If the Edge TPU delegate fails to load, it loads a non-TPU TFLite model or the universal SSD MobileNet CPU model, and reports device as cpu in the ready event.

Why does the Coral TPU stop returning detections after running for a while?▼

The TPU may be thermally throttling or the USB connection may have dropped. The built-in watchdog detects hangs with a 10-second invoke timeout and flags stalls after 30 consecutive empty frames, prompting the host to restart the process.

Related Skills

Explore More Software Engineering Skills→
stevengonsalvezstevengonsalvez

camsnap

Capture snapshots, clips, and motion events from RTSP/ONVIF cameras using ffmpeg.

Community
Intermediate
LuminaVaultLuminaVault

Market Alert Service

Monitor stocks, crypto, and news markets and send alerts to Discord, Telegram, or Slack.

Official
Advanced
autonomous-aiautonomous-ai
10

face-enroll

Manage the lamp's face recognition roster via HTTP endpoints.

Official
Advanced
dongg622dongg622

ppu-inference-yolo

Deploy YOLO object detection models on PPU hardware for real-time inference.

Community
Intermediate