Raspberry Pi just announced software updates for its growing suite of AI products. This is coming with full support for the AI HAT+ on the Trixie release of Raspberry Pi OS and a brand new debugging feature for the AI Camera.
The Raspberry Pi AI HAT+ and the corresponding AI Kit, which both rely on Hailo AI accelerators, are now fully compatible with the recently launched Trixie version of Raspberry Pi OS. You can get all the necessary software packages right from the standard apt repository and start building out your machine learning projects quickly.
The hardware itself is pretty powerful for a Pi add-on, especially when you look at the specifications. The AI HAT+ is rated for 26 TOPS, which immediately tells you this board is built for heavy lifting like real-time image processing and serious neural network acceleration.
There is a big technical change here that I think is actually a huge win for people who tinker with these modules frequently. Raspberry Pi decided to remove the Hailo device driver from the core kernel builds. Instead, the team will now use Dynamic Kernel Module Support (DKMS) to build and install that kernel driver during the package installation process.
This decoupling is a really smart move because it gives the development team more flexibility when they issue future software releases. Crucially, this change lets users downgrade the device driver without being forced to also downgrade the entire kernel.
If you previously generated custom-built models using an older version of the Hailo Dataflow Compiler, the ability to roll back just the driver is essential for maintaining compatibility with your existing work. The installation steps are almost exactly the same as before, but you now have to make sure you install the DKMS framework first before you install the main ‘hailo-all’ package.
The Raspberry Pi AI Camera also received a major update addressing a feature that developers have apparently been asking for since the camera launched. The team implemented an input tensor injection feature on the AI Camera. This tool lets you easily debug custom or purpose-built neural networks running directly on the device.
I would say the biggest problem when deploying custom AI models is verifying that they actually work correctly under various conditions after they are compiled. Input tensor injection solves that issue by providing a reliable testing methodology. This feature lets you validate the quality and performance of the network by repeatedly feeding it images from an existing dataset.
Whether you are using a standard, well-known dataset like COCO or something entirely custom and tailored to your specific application, the process makes sure you have reliable and repeatable testing. This is the kind of feature that dramatically improves the development experience for anyone serious about edge computing.
To grab this new debugging functionality for your AI Camera, you just need to run the standard update commands. You should run ‘sudo apt update’ and then follow up with ‘sudo apt full-upgrade -y’ to make sure your system is completely current. There is even an example input tensor injection script available if you want to try the feature out right away.
Source: Raspberry Pi
