CanaKit Raspberry Pi 4 8GB Starter Kit (Amazon Affiliate link) - [ Ссылка ]
During my last video [ Ссылка ] I asked you whether you are interested in image recognition software and you said - "Yes" big time, so I listen and execute. In this video I will show you how you can use #TensorFlow to perform real-time object detection using images streamed from the #Raspberry #Pi Camera.

Requirements
To complete this tutorial you will need:
Raspberry Pi with Raspbian OS - [ Ссылка ]
Connected and configured Pi Camera
Raspberry Pi Camera Module V2-8 Megapixel,1080p - [ Ссылка ]
Cheaper camera + Holder - [ Ссылка ]
Some kind of monitor connected to the Raspberry Pi to see the camera stream.
(really Optional) Coral USB Accelerator - [ Ссылка ]
And as usual smashed SUBSCRIBE button if you haven’t done that already.

What is TensorFlow?
It is an open source library to help you develop and train Machine Learning models. You can run TensorFlow locally on your devices and no info or images are sent to the cloud. So privacy concerned guys this is just for you.

How objects detection looks like?
You will see it in details at the end of the video, but basically when you put some objects in front of the Raspberry Pi camera you'll see color boxes drawn around objects that the model recognizes. The recognised objects will have a label and probability score. Higher score means that the model is pretty sure that the object is what is written on the label.

TensorFlow Lite vs Tensorflow
We are going to install TensorFlow Lite which is much smaller package than TensorFlow.

I will test this on my Raspberry Pi 3, if you have Pi 4 it will run even better.

So, Without further ado lets install this TensorFlow lite on a Raspberry Pi and start to classify images:

Steps:
* To enable Raspberry Pi camera type the following in the Raspberry terminal:
sudo raspi-config
* then go to Interfacing Options\Camera and on the dialog choose 'Yes'
* To test the camera type:
raspistill -o ~/Desktop/like_this_video.jpg
* You can then open the ~/Desktop/like_this_video.jpg file to check that you have a still image from your Pi camera
* If everything is OK with your camera, check your installed python version by:
python3 —version
* Then go to the TensorFlow python quickstart - [ Ссылка ]
* Copy the URL matching to your python version and OS and put it after pip3 install. example: pip3 install YOUR_URL.
* If you have a python version 3.7.x you can use the following command:
pip3 install [ Ссылка ]
* Clone TensorFlow examples repo to your Raspberry Pi like this:
git clone [ Ссылка ] --depth 1
* Enter inside this folder with the the command below:
cd examples/lite/examples/object_detection/raspberry_pi
* Then use the script below to install a couple Python packages, and download the MobileNet model and labels file. The script takes an argument specifying where you want to save the model files
bash download.sh /tmp
* Run the example with this:
python3 detect_picamera.py \
--model /tmp/detect.tflite \
--labels /tmp/coco_labels.txt

Congratulations, you should see the camera feed appear on the monitor attached to your Raspberry Pi.

The Goal
The ultimate goal for me is to make this TensorFlow Lite image classification to work with any RTSP camera and of course with Home Assistant and Node-RED.

So you will be able to receive notifications only when there is motion detected by person and not a cat or wind for example. That will be really nice feature to any Smart Home security system.

Question?
It is possible to build and train your own Machine Learning models with TensorFlow. What will be most beneficial for you to be trained? Let me know in the comments section below.

Coral USB Accelerator
It is possible to accelerate the image classification by 10X by using the Coral USB Accelerator - [ Ссылка ] This is a small usb drive capable of performing 4 trillion operation per seconds with very low consumption power and it is compatible with TensorFlow and Raspberry Pi.

Unfortunately I don’t have such device, so I cannot show you this at the moment. But If I manage to find one I will create a follow-up video.

I really hope that you find this information useful and you now can detect objects on your Raspberry Pi using TensorFlow lite.

Thank you for watching, stay safe and see you next time.

DISCLAIMER
Some of the links above are affiliate links, where I earn a small commission if you click on the link and purchase an item. You are not obligated to do so, but it does help fund these videos in hopes of bringing value to you!

свернуть