Deep learning-based spacecraft detection algorithm optimized for embedded hardware

The problem of space debris is of great concern to the international scientific community, as it might preclude access to space for future missions and negatively influence existing ones. For this reason, aerospace companies worldwide are actively developing on-orbit servicing and space debris removal missions. In order for these missions to be successful, highly autonomous navigation systems are essential. To enable accurate and reliable vision-based navigation, robust object detection capabilities are paramount. Deep Learning (DL) algorithms are increasingly recognized for their potential to achieve the required level of autonomy and accuracy in this critical task. This thesis focuses on advancing the application of DL for spacecraft object detection by developing and optimizing an image classification network. Specifically, we leveraged the EfficientNet-B0 architecture to create a lightweight yet accurate object detection network tailored for deployment on embedded hardware, such as the Jetson Orin Nano. To further enhance efficiency, we explored quantization techniques, converting the model through Torch, ONNX, and TensorRT formats. Additionally, we trained a YOLO network to provide a comparative benchmark for our developed solution. Both networks were trained using the SPEED+ dataset, a comprehensive resource for spacecraft pose estimation. The primary aim of this work is to demonstrate the feasibility of deploying highly efficient DL-based object detection on resource-constrained platforms, thereby contributing to the development of robust and autonomous space debris removal missions. Indeed, our software pipeline was capable of running at up to 136.05 fps showing satisfactory object detection performance, with an average Intersection over Union index equal to 0.850 on synthetic images, which drops to about 0.382 during domain gap tests.