InSegNet: Semantic segmentation-based lane detection network for clear and occluded roads

The global surge of vehicles has resulted in a significant increase in traffic accidents. Consequently, the integration of Advanced Driver Assistance Systems (ADAS) in vehicles is essential to ensure road safety. Existing lane detection systems are still unable to accurately detect curvy roads, especially at night. To address this limitation, we propose a novel and robust lane detection network, InSegNet, based on a semantic segmentation algorithm. A total of four networks, named InSegNet-v1, InSegNet-v2, InSegNet-v3, and InSegNet-v4, are designed, where the encoder comprises two convolutional stages followed by two inception stages for strong feature extraction. InSegNet-v1, InSegNet-v2, and InSegNet-v3 follow the kernel pattern of GoogleNet, while InSegNet-v4 adopts the kernel strategy of the SegNet semantic segmentation framework. InSegNet-v2, InSegNet-v3, and InSegNet-v4 include additional connections between the encoder and decoder to enhance feature propagation. Furthermore, InSegNet-v3 and InSegNet-v4 incorporate residual connections in the encoder to improve feature extraction and network performance. The networks are trained and tested on a self-collected dataset consisting of clear and occluded road images captured during both day and night on local roads. The performance of all networks is evaluated by F1-score, Jaccard Index, True Positive Rate (TPR), False Positive Rate (FPR), and Precision, and compared with benchmark networks. InSegNet-v3 achieves a 1.4% higher TPR than ENet and a 2.54% improvement in Precision over ERFNet. Whereas InSegNet-v4 demonstrates a 25.6% reduction in FPR and a 1.32% increase in Precision compared to ERFNet. Based on superior performance across key evaluation metrics, our networks are well-suited for deployment in intelligent vehicles, particularly in scenarios requiring higher accuracy, such as nighttime driving and curvy roads.