OpenCV

The Open Computer Vision Library (OpenCV) [6] has another set of libraries that you can use on your Raspberry Pi. OpenCV is used for gesture, face, and object recognition and classification. The OpenCV deep neural network (DNN) module works with pre-trained networks for this purpose and can be used in combination with TensorFlow Lite. To install OpenCV on the Raspberry Pi, though, you need to resolve a large number of dependencies, and you need to specify manually a large number of flags during the build. This difficulty prompted Dutch AI specialists at Q-engineering [7] to publish a freely available and BSD-licensed script on GitHub that lets you work around these steps. To install and run this OpenCV script, enter:

$ wget https://github.com/Qengineering/Install-OpenCV-Raspberry-Pi-64-bits/raw/main/OpenCV-4-5-5.sh
$ sudo chmod 755 ./OpenCV-4-5-5.sh
$ ./OpenCV-4-5-5.sh

As a final step, you need to integrate the graphical Code::Blocks integrated development environment (IDE) [8] into your system (Figure 2). With its help, you can then use TensorFlow Lite and OpenCV to recognize and classify objects by drawing on various sample networks. These capabilities apply not only to photos, but also to livestreams from the connected camera. Code::Blocks supports the C and C++ programming languages and is therefore ideally suited for AI applications. The command

sudo apt-get install codeblocks

Figure 2: The Code::Blocks IDE helps you use AI models.

installs the package and automatically creates a starter on the desktop and in the Raspberry Pi OS menu system.

Examples

After completing the installation, you can test some sample scenarios by drawing on a number of prefabricated and trained code examples from Q-engineering; all of these achieve very good results on the Raspberry Pi 4, even in livestreams [9]. Code::Blocks is used here, too, and it even provides slide shows of screenshots in the tutorials to help newcomers gain some initial experience with AI applications [10]. Instead of the sample photos and MP4 videos included in the bundle, you can use your own pictures or video files from the Raspberry Pi camera. All you need to do is copy them to the appropriate directories and specify them as parameters in Code::Blocks (Figure 3).

Figure 3: The object recognition elements are shown in the original image along with the percent likelihood of correct recognition.

Generating Your Models

Because custom models cannot be trained on small computers, Google offers a web-based tool [11] to help in the creation of models. The tool is suitable for various model types and outputs them as files in the TensorFlow format so that you can use the models in the Lite variant after converting. Please note, however, that generating a model for object recognition (e.g., on images and photographs) means uploading several hundred sample images. The sample images also need to be high resolution to achieve high accuracy levels later. You need to schedule several hours to work with the tool (Figure 4).

Figure 4: You can create your own models with a web-based tool.