Constraint-driven neural network architecture adaptation for image localisation on a miniature mobile robotPleterski, Jan (Avtor) Puc, Jernej (Avtor) Vrabič, Rok (Avtor) Podržaj, Primož (Avtor) Škulj, Gašper (Avtor) convolutional neural networklow-powermicrocontrollerminiature mobile robotMobileNetobject localisationtiny machine learningTinyMLDeploying deep learning models on low-power devices is an increasingly important research challenge. Microcontrollers, in particular, have strict limitations due to the limited processing power and memory. Specifically, object localisation on such devices remains in the early stages of research, showing the need for architectural designs tailored to resource-constrained environments. In this work, we implement and evaluate different architectural configurations of deep learning models on a miniature mobile robot operating in a granular, sand-like terrain, where both robots and rocks must be localised as a prerequisite for robust navigation. The localisation models are based on the MobileNetV2 architecture utilising the FOMO (Faster Objects, More Objects) algorithm with dynamic convolutions. The proposed system consists of an HM01B0 greyscale camera integrated with a low-power RP2040 microcontroller. The object detection models are trained on a custom dataset consisting of 10,000 greyscale images (160×160 px) featuring robots and rocks. Our findings indicate that the best trade-offs between accuracy and memory are achieved by truncating the middle layers of the base model, balancing output resolution with retained pre-trained features and classifier complexity. Additionally, the results indicate that increasing the input image resolution produces a greater improvement in accuracy compared to expanding the model’s classifier head, relative to the increase in inference time. This study presents the first large-scale variation of FOMO truncations for the purpose of object localisation on low-power microcontrollers and offers practical guidance for selecting neural network architectures.20262026-05-08 09:14:04Članek v reviji182360sl