An instance segmentation algorithm was adopted and a lane detection method based on improved hybrid task cascade (HTC) network was proposed aiming at the problems of fuzzy feature representation and low semantic information utilization in the lane detection method based on semantic segmentation. Deformable convolution was introduced into the backbone network based on HTC network model in order to improve the ability of the backbone network to extract lane features in complex environments. The structure of feature pyramid network was improved by adding a bottom-up low-level feature transmission path based on feature pyramid network, and dilated convolution was introduced to increase the receptive field of feature map without loss of lane feature information. The accurate location information of lane lines contained in the low-level features was used to improve the accuracy of lane detection. The experimental results show that the improved HTC network model can realize the robust extraction of lane features, obtain better detection performance in complex road environments, and effectively improve the accuracy of lane detection.