To address the problems of slow prediction speed and low segmentation accuracy of existing semantic segmentation methods for highway guardrail detection, an UAV highway guardrail detection method based on improved DeepLabV3+ is proposed. First, the MobileNetv2 network was used to replace the backbone of the original model and outputs the middle layer's features to reduce number of parameters while recovering the spatial information lost in the downsampling process; then an atrous spatial pyramid pooling was improved by the densely connected expansive convolution to reduce the phenomenon of missed segmentation; finally, the spatial group-wise enhance (SGE) attention mechanism was introduced in the encoder part to reduce the phenomenon of wrong segmentation. The experimental results show that the average intersection over union, average pixel accuracy, and frames per second transmission of the improved model can reach 79.20%, 87.89%, and 52.59, which are 2.59%, 2.93%, and 56.70% higher than the base model, respectively, and number of parameters is reduced by 78.85%. This method can thus improve the segmentation accuracy for the guardrail while guaranteeing the model's prediction speed.