Implementation of Road Segmentation Using U-Net Model on Single Board Computer

Technological developments in the era of globalization, several companies are competing in the field of artificial intelligence by developing autonomous drive systems. Training and road segmentation testing in this study were carried out using deep learning with the U-Net architecture method. The advantage of this method over other methods is that U-Net retains the full context of the input image. The road segmentation algorithm is implemented using Python programming. The algorithm is then executed on a single board computer. The parameters sought are the first accuracy with a value of 91.65 %, the second precision with a value of 75.53 %, the third recall with a value of 93.19 %, the fourth F1-Score with a value of 82.99 %, and the last IOU with a value of 71.36 %. The live segmentation algorithm can still detect roads, but some scenes involving objects are not included in the road segmentation process. This is due to several factors such as confusing light conditions, blur, shadows, and colors. Live segmentation using RGB frame produces an average FPS value of 0.30, and without RGB produces an average FPS value of 1.82.