Semantic Segmentation Method for Drivable Areas under Unstructured Road

To improve the ability of autonomous vehicles to recognize drivable areas on park roads, this study puts foward a semantic segmentation approach based on the STDC (Short-Term Dense Connection) network. On this foundation of the STDC network, a deep convolutional neural network with an encoder-decoder structure is designed, and a double-chain cascaded feature fusion module (DCFFM) is introduced. Additionally, the Convolutional Block Attention Module (CBAM) is incorporated to dynamically adjust channel and spatial attention, thereby reinforcing the model s capacity to discern crucial features. A dataset suitable for drivable area segmentation in park environments was collected and used for training and validation. The experimental outcomes demonstrate that the algorithm attained a mIoU of 98.41% on the self-built dataset, representing a 2.73 percentage point improvement over the STDC network. Furthermore, the mIoU on public datasets exhibited a 2.55 percentage point enhancement in comparison to that of the STDC network.