Abstract: In the scenes of object occlusion, lighting changes and shadow interference, the detection of the point instance network（PINet） has high accuracy but with poor real-time performance. A lane detection model combining the recurrent feature-shift aggregator（RESA） algorithm was proposed to ensure the accuracy of the PINet model and improve the inference speed of the network. To eliminate the redundant multi-scale operations and accelerate the inference speed of the model, only one bottle-neck network was used for the prediction network model after the computational power analysis. To compensate the accuracy decreasing by module pruning, the RESA module was proposed to capture the spatial information across rows and columns in the image for enhancing the lane line features. The tests of the improved model were conducted on the Tusimple, CULane and Custom datasets. The results show that the improved network model performs well in various complex scenarios of object occlusion, lighting changes and shadow interference with improved accuracy and real-time processing speed of lane segmentation. The detection and recognition performance is better than that of traditional PINet network algorithms. Except for slight improvement in the F1 indicator, the inference speeds of the improved model are respectively increased by 20.3%, 52.9% and 13.9% in the three datasets.