Implementasi Semantic Segmentation menggunakan Arsitektur U-Net pada Robot KRSRI untuk Memahami Lingkungan Arena

Abstract

Autonomous robots in the Indonesian SAR Robot Contest (KRSRI) require detailed scene understanding capabilities to navigate effectively in complex and dynamic arenas. Traditional feature-based or bounding box-based detection methods have limitations in providing spatial information regarding object shapes and boundaries. This research implements a semantic segmentation method using the U-Net architecture to perform pixel-wise classification of 11 object classes within the arena environment. The model was trained on a dataset of 355 original images, augmented to 1065 images, and subsequently deployed on a Raspberry Pi 5 in TensorFlow Lite format for real-time testing. The evaluation on the test set demonstrates the model's excellent performance, achieving a mean Intersection over Union (mIoU) of 0.8253 and a Dice Score of 0.9004. When implemented on the Raspberry Pi 5, the model runs at an average inference speed of 2.8 FPS or around 357 ms per frame using the CPU, proving viable for lightweight scene understanding needs. This research demonstrates that the U-Net architecture can be successfully applied to provide detailed and accurate visual perception for the KRSRI robot on an edge device.