Pengembangan Model Segmentasi Citra untuk Autonomous Wheelchair Menggunakan Teknik Knowledge Distillation di Embedded System

Abstract

High-performance segmentation models for autonomous wheelchairs, such as ResUNetFormer, face significant challenges in real-time deployment on embedded systems due to high computational complexity and excessive memory consumption. This study aims to develop efficient image segmentation models by applying Knowledge Distillation (KD) to transfer knowledge from a complex teacher model to lightweight student models. Using a quantitative experimental approach, ResUNetFormer served as the teacher model, while BiSeNetV2 and Fast-SCNN were selected as student models. The methodology integrated pixel wise and pair-wise distillation losses to optimize performance on a dataset consisting of 1,500 corridor images. Models were deployed and evaluated on an NVIDIA Jetson Orin NX based on mIoU, frames per second (FPS), and memory usage. Results demonstrate that BiSeNetV2 with KD achieved the best student performance with a mIoU of 84.55% and an inference speed of 35.24 FPS. This model significantly reduced memory usage to 1.5 GB compared to the teacher's 5.1 GB. Meanwhile, Fast-SCNN with KD reached 46.4 FPS with a mIoU of 82.61%. Overall, the application of KD improved student mIoU by 1.08% to 2% without increasing the number of parameters. In conclusion, Knowledge Distillation effectively produces high-accuracy, lightweight models suitable for real-time edge deployment on autonomous wheelchairs, successfully overcoming the limitations of large-scale architectures.