AI on a Pi? Believe it!
AI Summary
- Pineberry AI Hat Overview
- Connects to Raspberry Pi 5 via PCIe express bus
- Features M2 slot for Coral AI Edge TPU
- 2,000 CPU
- Interface supports gen 3 speeds for enhanced AI performance
- Frigate NVR Home Surveillance Test
- Raspberry Pi setup with Edge TPU and $15 webcam
- Surveillance records upon motion detection
- Achieved faster inference times than recommended hardware
- Setup Process
- Mount TPU onto AI hat with spacers and screws
- Secure 16p FPC ribbon
- Connect AI hat to 8 GB Raspberry Pi 5
- Total setup cost: 19 for AI hat, $25 for Coral AI chip)
- Performance and Configuration
- PCIe version offers better thermal management than USB version
- Raspberry Pi OS 64bit setup with micro SD card, Wi-Fi, and SSH
- Script to install drivers and configure operating system
- Testing the TPU
- Verify TPU installation and expose to Docker
- Use pyCoral library and Frigate for home surveillance NVR
- Docker Debian 10 VM setup for running pyCoral
- Install MQTT for messaging relay service
- Configure Frigate with YAML file and IP address
- Frigate Optimization and Usage
- Add FFmpeg preset for Raspberry Pi 64bit h264 hardware acceleration
- Set up web UI for camera management
- Person detection records events when a person is in frame
- Performance Insights
- Smooth operation on Raspberry Pi 5
- 7.49 millisecond inference time
- Potential Enhancements and Alternatives
- PCIe may offer lower latency than USB
- Camera Module 3 with 12 MP sensor for HD IoT camera
- Possibility of dual Edge TPU setup
- Raspberry Pi may release an official M2 hat for NVMe storage
- Single TPU can support around 10 cameras
- Personal Recommendations
- Consider USB accelerator to keep PCIe slot open for NVMe storage
- Explore custom TensorFlow Lite models with pyCoral library
For more details, refer to the full blog post in the video description.