{"id":41336,"date":"2021-03-22T11:01:11","date_gmt":"2021-03-22T03:01:11","guid":{"rendered":"\/blog\/?p=41336"},"modified":"2021-03-22T11:01:15","modified_gmt":"2021-03-22T03:01:15","slug":"machine-learning-powered-inventory-tracking-with-raspberry-pi","status":"publish","type":"post","link":"https:\/\/www.seeedstudio.com\/blog\/2021\/03\/22\/machine-learning-powered-inventory-tracking-with-raspberry-pi\/","title":{"rendered":"Machine Learning Powered Inventory Tracking with Raspberry Pi"},"content":{"rendered":"\n

In today’s tutorial, I will show you how to create a smart inventory tracker using object detection, powered by deep learning, with just a Raspberry Pi 4 and a camera. We will apply transfer learning on the YOLOv4 tiny model to identify custom objects, then use a simple python script to parse the model’s output to produce a count of each object. Finally, we will also integrate the application with Azure IoT Central so that we can monitor our inventory remotely and conveniently.<\/p>\n\n\n\n

<\/div>\n\n\n\n
\"\"\/<\/figure><\/div>\n\n\n\n
<\/div>\n\n\n\n
\n\n\n\n
<\/div>\n\n\n\n

Project Overview<\/strong><\/h2>\n\n\n\n
<\/div>\n\n\n\n

Have you ever passed by a grocery store but found yourself unsure of whether you needed to get that extra carton of milk? Well, what if there was some way to have eyes on the inside of our fridge to update us with that information? Today, with machine learning and IoT infrastructures, we are going to turn that convenience into reality.<\/p>\n\n\n\n

To summarize, we want to create two key features in our project. First, to automatically count the items in our fridge. Then, to be able to access the data remotely when we need it. Before we begin, let’s take a look at a short video demonstration below.<\/p>\n\n\n\n

<\/div>\n\n\n\n
\"\"<\/figure><\/div>\n\n\n\n
<\/div>\n\n\n\n

Machine Learning: Image Classification vs Object Detection<\/strong><\/h2>\n\n\n\n
<\/div>\n\n\n\n

Object recognition tasks like image classification and object detection are both performed with deep neural networks, but it is meaningful to differentiate between the two. In image classification, our input is an image that contains a single class. For example, we might have a photo of a single cat, or a dog, that we want the model to classify as a single output.<\/p>\n\n\n\n

In object detection, however, we want the model to not only identify multiple objects, but to also locate all the instances of each class present in a photo. Indeed, as opposed to classification, we will now be able to count the number of each type of object present in the frame – just what we need for this project!<\/p>\n\n\n\n

\"\"
Source: DataCamp<\/em><\/figcaption><\/figure><\/div>\n\n\n\n
<\/div>\n\n\n\n

For more information, Jason Brownlee’s article<\/a> is a great introduction to the different types of computer vision tasks.<\/p>\n\n\n\n

<\/div>\n\n\n\n

What is Transfer Learning?<\/strong><\/h2>\n\n\n\n
<\/div>\n\n\n\n

Because computer vision tasks in general are considered complex tasks, it can take days to train a functional model even on a powerful GPU. Fortunately, as long as we have both the architecture and the weights, anyone can easily implement existing models that have already been trained.<\/p>\n\n\n\n

However, what if we want to train the model to recognise other things? After all, it’d be a farfetch to expect these models to naturally be great at recognising our household’s specific choice of grocery products. In this case, we don’t have to train a new model from scratch. Instead, we can apply transfer learning.<\/p>\n\n\n\n

<\/div>\n\n\n\n
\"\"
Source: datascience.aero<\/a><\/em><\/figcaption><\/figure><\/div>\n\n\n\n
<\/div>\n\n\n\n

Transfer learning is a technique where we utilise a pretrained model, and train it with a new dataset with the objective of accomplishing a similar task. It allows us to achieve higher performance in our models with less data, while still drastically reducing the amount of time and resources required for training. Today, we will perform transfer learning with YOLOv4 Tiny.<\/p>\n\n\n\n

<\/div>\n\n\n\n

What is YOLO?<\/strong><\/h2>\n\n\n\n
<\/div>\n\n\n\n

YOLO, which stands for You Only Look Once, is a state of the art neural network machine learning framework for object detection. When its first paper was released in 2015, it beat out the other models in inferencing speed, allowing for real time object detection at some expense of accuracy. Today, work on YOLO has continued under different parties to bring improvements in accuracy while retaining low latency.<\/p>\n\n\n\n

For today’s project, we are using the 4th version, YOLOv4<\/a>. More specifically, we will be using the tiny version of YOLOv4, which has a smaller network size. While we do sacrifice some accuracy with a smaller model, this brings a reduction in the computing power requirements, allowing the model to be run with decent performance on an edge device like our Raspberry Pi.<\/p>\n\n\n\n

<\/div>\n\n\n\n
\n\n\n\n
<\/div>\n\n\n\n

Prepare Your Custom Dataset<\/strong><\/h2>\n\n\n\n
<\/div>\n\n\n\n

As our first step to train our custom model, we will have to build our dataset by gathering photos of the items that we want to detect. You’ll want at least 50 images of each class. In my project, I decided to count the number of milk cartons, eggs, and yoghurt packets, so my dataset should end up with at least 150 images in total.<\/p>\n\n\n\n

<\/div>\n\n\n\n

Here are some tips when collecting your dataset.<\/strong><\/p>\n\n\n\n