RPI Vision

Deep object detection on a Raspberry Pi using Tensorflow & Keras.

Materials

• Raspberry Pi 3 Model B
• SD card 8+ GB
• 3.5" 480 x 320 TFT/SPI screen (XPT2046 controller)

Install Dependencies (on Raspberry Pi)

• Install Raspbian
• Configure WiFi (optional, but recommended)
• @todo link to other basic Pi configuration tasks outside the scope of this guide (add authorized SSH keys, disable password, change hostname)
• Install system dependencies

sudo apt-get update && \
sudo apt-get upgrade && \
sudo apt-get install git python3-dev python3-pip \
crossbuild-essential-armhf libatlas-base-dev   \
libhdf5-dev libhdf5-serial-dev \ 
libopenjp2-7-dev ibtiff5 build-essential cmake pkg-config && \
sudo pip3 install -U virtualenv

git clone [email protected]:leigh-johnson/rpi-vision.git
cd rpi-vision
pip install -r rpi.requirements.txt

Install TFT Drivers

WARNING these instructions only apply to the 3.5" TFT (XPT2046) screen. If you're using a difference size or controller, please refer to the instructions in LCD-show#README.

git clone [email protected]:goodtft/LCD-show.git
chmod -R 755 LCD-show
cd LCD-show
sudo ./LCD35-show

Install FBCP

This step is only neccessary if you're using an SPI Display. If you're using an HDMI display, skip this step.

Updating /boot/config.txt

For better TFT screen performance, add the following to /boot/config.txt. Refer to Raspbian's video options in config.txt if you're using a different display.

@ todo

Setup Google Cloud (optional)

@todo

Running a trainer (GPU Accelerated)

pip install -r trainer.requirements.txt

@todo API docs

Training a custom CNN

@todo API docs

Analyzing via Tensorboard

tensorboard --logdir gs://my-gcs-bucket/my-model/logs/

References

• Training a neural network from little data
• How to easily Detect Objects with Neural Nets
• d4, d6, d8. d10, d20 images