BiSeNet-CBAM: An Enhanced Framework for Real-Time Semantic Segmentation of Unstructured Scenes with Dual Loss Optimization

This paper presents an enhanced Deep Learning framework for real-time semantic segmentation of unstructured road scenes, extending BiSeNetV1 with the Convolutional Block Attention Module (CBAM) to emphasize relevant features and improve segmentation performance. A hybrid loss function combining cross-entropy and IoU loss improves boundary accuracy while maintaining efficiency. Benchmarking on the Indian Driving Dataset (IDD) demonstrates superior real-time performance, achieving a mean Intersection over Union (mIoU) of 66.76% and an FPS of 518.97—over ten times faster than the baseline. Furthermore, the proposed model also achieves superior performance on the CamVid dataset, with an mIoU of 48.82%, outperforming all tested real-time models. These results highlight the potential of attention mechanisms and lightweight architectures in advancing autonomous driving and unstructured traffic analysis.