This project developed a real-time semantic segmentation model for an autonomous robot to identify and follow outdoor trails across varying terrain types, including dirt, gravel, asphalt, and concrete. A custom dataset of 1,000 manually labeled images was created to provide pixel-level trail annotations for training and evaluation. Several state-of-the-art segmentation models (LEDNet, DeepLabV3, ICNet, BiSeNet, and PSPNet) were implemented in PyTorch and compared using pixel accuracy, mean Intersection-over-Union (mIoU), and runtime performance. LEDNet with a ResNet-50 backbone was selected for its balance of accuracy and speed, and its performance was further optimized by tuning key hyperparameters such as learning rate, batch size, and number of training epochs.

More information on the integration of this model with the physical robot may be found here.