The lane detection model based on a single feature layer is plagued by an imbalance in lane recognition and positioning, particularly when lanes are occluded or lighting conditions are poor. This significantly hampers the enhancement of environment understanding capabilities in driving assistance or autonomous driving systems. To address this challenge, a hybrid attention lane detection model based on multi-feature scale-by-scale regression, named HAMFD, is proposed. The model first employs a hybrid attention mechanism to capture lane features that encompass both global and local contextual information, thereby enhancing lane representation and improving adaptability to occluded lanes and poor lighting conditions. It then leverages scale information from different feature layers to regress lanes based on shape anchor information, progressing from larger to smaller scales to improve lane positioning accuracy. To enhance the quality and generalization ability of shape anchors, they are incorporated as trainable parameters for continuous updating and optimization during model training. The local tilt angle is also introduced into the IOU calculation to impose stricter shape constraints on lanes. Finally, a multi-task learning mechanism is adopted to adaptively learn the weights of each loss function, reducing the complexity of model parameter tuning. Tests demonstrate that the model can effectively detect lanes in complex real-world road scenarios. When equipped with a DLA34 backbone network, the model achieves an F1 score of 80.21% on the CULane dataset, with a detection rate of 128 fp/s, outperforming other lane detection models such as CondLaneNet, GANet, and Lane2Seq.