Semantic Segmentation Based Lightweight Lane Detection Network (LW Net) for Intelligent Vehicles

A novel lane detection system is proposed for intelligent vehicles. A key feature of this system is its lightweight design, which requires less computational power. Our lightweight network (LW Net) for semantic segmentation comprises convolutional and separable convolutional layers. We designed a total of six lightweight encoder models (LW Net-A, LW Net-B, LW Net-C, LW Net-D, LW Net-E, and LW Net-F), each paired with matching decoders. The first group of three models is based on depth D1, while the remainingmodels are based on depth D2. In these models, convolutional layers are either fully or partially replaced by separable convolutional layers. The lightweight network LW Net-A achieved an 88% reduction in training parameters, along with a 2.45% increase intest accuracy compared to the benchmark SegNet model. Meanwhile, LW Net-F attained a 2% increase in test accuracy and a remarkable 94% reduction in training parameters compared to the benchmark SegNet model. Overall, the proposed models are less computationally demanding than other benchmark networks, without compromising the pixel accuracy of the semantic model.