| Abstract Scope |
We present a comparative evaluation of U-Net, FCN-ResNet50, and DeepLabV3-ResNet50, for outdoor autonomous mobile robot navigation. The experiments were conducted using an outdoor mobile robot segmentation dataset containing RGB images and pixel-level segmentation masks with four semantic classes: Road, Sidewalk, Cycle Path and Other. The images and masks were resized to 192×512 pixels, and all models were trained for 200 epochs in high-performance computer. U-Net was implemented as a custom encoder-decoder segmentation architecture, while FCN-ResNet50 and DeepLabV3-ResNet50 used residual backbones for dense pixel-wise prediction. The models were evaluated using precision, recall, F1-score, Dice score, IoU, class accuracy, mean IoU, IoU threshold metrics, and computational cost. Based on the class-wise full-test evaluation, DeepLabV3-ResNet50 achieved the highest mIoU of 0.9147, followed by FCN-ResNet50 with 0.9059 and U-Net with 0.8362. These findings indicate that residual backbone-based segmentation models provide stronger and more consistent performance for outdoor AMR scene understanding. |