Real-Time AI-Based Image Processing Architectures for Lower-Limb Exoskeleton Control
Author
Chaudhry, Muhammad Hashim
Term
4. term
Publication year
2026
Submitted on
2026-06-04
Pages
107
Abstract
This thesis explores how a camera can be used in real time to tell whether stairs go up or down, to help control a lower-limb exoskeleton. An exoskeleton is a wearable robotic device that assists movement, and the system runs on small, low-power edge hardware. A wrong classification at a staircase could trigger an incorrect mode. We use a binary classifier that decides between ascending and descending stairs from a forward-facing RGB camera. Three lightweight architectures (ResNet-18, MobileNet v3 Small, and YOLOv8 nano-cls) were trained from scratch under one matched setup on the ExoNet stair subset, exported to ONNX (a standard format), and benchmarked on a Raspberry Pi 4 for accuracy, latency (time per decision), energy use, and model size. The accuracies are close: MobileNet 95.96%, ResNet-18 95.93%, YOLOv8 94.23%. MobileNet v3 Small is recommended for its tied-best accuracy, fastest inference (20.89 ms), lowest energy (124 mJ), and smallest model (6.09 MB). An accompanying business and regulatory analysis shows the technical and commercial recommendations align.
Dette speciale undersøger, hvordan et kamera kan bruges i realtid til at skelne mellem op- og nedadgående trapper for at styre et exoskelet til benene. Et exoskelet er en bærbar robot, der hjælper med bevægelse, og systemet kører på små, strømbesparende enheder (edge-hardware). Hvis overgangen ved en trappe bliver klassificeret forkert, kan exoskelettet vælge en forkert tilstand. Vi bruger en binær klassifikator, der afgør op eller ned ud fra et fremadrettet RGB-farvekamera. Tre letvægts-arkitekturer (ResNet-18, MobileNet v3 Small og YOLOv8 nano-cls) blev trænet fra bunden med samme opskrift på ExoNets trappe-delmængde, eksporteret til ONNX (et standardformat), og testet på en Raspberry Pi 4 for nøjagtighed, latenstid (tid pr. beslutning), energiforbrug og modelstørrelse. Nøjagtighederne ligger tæt: MobileNet 95,96%, ResNet-18 95,93%, YOLOv8 94,23%. MobileNet v3 Small anbefales, fordi den har delt topnøjagtighed, hurtigste inferens (20,89 ms), lavest energi (124 mJ) og den mindste model (6,09 MB). En tilhørende forretnings- og regulatorisk analyse viser, at de tekniske og kommercielle anbefalinger stemmer overens.
[This apstract has been rewritten with the help of AI based on the project's original abstract]