You can read more about our CanSat mission, which involves developing a miniature satellite housed within the dimensions of a standard soda can, by checking in for regular project updates here.
Mission essentials
CanSat is an educational initiative that gives students hands-on experience in designing, building, and testing a miniaturized pseudo-satellite with real-world applications. The primary mission, a key requirement for all CanSat projects, involves measuring atmospheric data — specifically temperature and pressure — during its descent from a launch vehicle.
Secondary mission
Our team came up with the idea to build stereo cameras into our CanSat design for our secondary mission. We plan to use data from both cameras to perform long-range depth estimation, which is crucial for tasks such as landing site selection and terrain mapping. Our initial research findings showed that traditional methods of depth estimation wouldn't measure up given the dynamic and unpredictable situation our can will find itself in. Factors such as varying light levels, vertical distance, and complex ground textures further complicate the process, which led us to look to machine learning (ML) for a solution.
A new take on depth estimation
Depth estimation using stereo cameras - two cameras working in tandem - is not straightforward. It usually involves comparing two images from slightly different points of view to determine the 3D structure of the scene. But, traditional computer vision algorithms would struggle with the complexity and noise present in the real-world data collected during our CanSat's descent.
Training neural network
This is where machine learning comes in. By training a neural network with a custom dataset of stereo image pairs and corresponding depth information, we can teach our CanSat to infer depth maps from the images it captures as it falls. The ML model can adapt to unpredictable contexts in a way that traditional algorithms can't.
Liftoff: design and build
Creating the physical CanSat posed its own set of challenges as the design brief is very strict. Everything must fit within the cylindrical constraints of a standard soda can, and also stand up to the durability requirements. We decided to 3D print our can to meet CanSat's design spec and fit our specialised hardware choices. The internal structure of our can has been designed to withstand the forces of ejection and impact upon landing despite the inherent strength limitations of 3D printed parts.
One of the critical considerations was ensuring the stereo cameras' alignment and stabilization, as accurate depth estimation hinges on precision. We created a mounting solution that holds the cameras in place and folds out as the can falls, while maximising the distance between the cameras, as this gives more accurate results from the depth estimation algorithm.
Demistifying software and ML
Our journey isn't just about showcasing skills. In the spirit of learning, we'll be rolling out a series of interactive articles to explain the hardware design, software and machine learning concepts driving our CanSat's capabilities.
These articles will offer clear and simple explanations and illustrative diagrams to help readers better understand the technology and methodologies at the core of our work. Whether you're new to this field or looking to expand your technical understanding, we hope you'll find value in these resources, as well as enjoying following our story.
The visualisation has a 3D element, and the view can be moved using your mouse or by swiping.
Hover over or tap the info marker to see a more detailed explanation of what the interactive element represents.
Next steps
Stay tuned for more updates as we continue to refine our technology, test our hypotheses, and ultimately launch our creation skyward. Thanks for following our journey and do stay connected with our blog for a series of upcoming posts where we'll delve deeper into the specifics of our CanSat's design, the intricacies of machine learning models, and the exciting outcomes of our project. The countdown to launch has begun!