Drivable area recognition on unstructured roads for autonomous vehicles using an optimized bilateral neural network

The accuracy of drivable area recognition is key to ensuring the safety of autonomous vehicles on unstructured roads. The diverse road shapes and ambiguous boundaries pose significant challenges for recognition tasks. Existing models often struggle to balance real-time performance and accuracy on complex roads. This study proposes a lightweight bilateral semantic segmentation network with a dual attention mechanism, capturing both channel and spatial information to handle unclear object boundaries and complex backgrounds. We integrate this mechanism into the BiSeNet framework's dual-path structure. The Efficient Channel Attention (ECA) mechanism is incorporated into the spatial path to extract spatial information critical for unstructured roads, while the Coordinate Attention (CA) mechanism is embedded in the context path to capture positional information. A residual network is employed in the context path to enhance efficiency and maintain a lightweight design, making the model suitable for autonomous vehicle systems. Additionally, a global convolutional network (GCN) and boundary refinement (BR) module further improve segmentation accuracy. Finally, the proposed model achieved 93.89% MIoU and 97.32% PA on the ORFD dataset with a speed of 62.49 FPS. Test results demonstrate that the model improves accuracy while ensuring real-time performance, outperforming other advanced real-time segmentation models. The proposed model offers a promising solution to the challenge of real-time, high-accuracy drivable area recognition, which is crucial for the safe and efficient operation of autonomous vehicles in dynamic and unstructured environments.