An obstacle detection model AO-YOLO was proposed in order to solve the problems of low accuracy and excessive calculation of obstacle recognition in traditional methods faced by general aviation aircraft in plateau mountain application scenarios. A multi-scale expanded attention mechanism was introduced by optimizing the existing YOLOv8n model in order to enhance the ability of Neck network to fuse features of different scales. HGBlock structure was used to replace the original Bottleneck module, and lightweight convolution was used to establish the feature hierarchical relationship, so that the network can extract local and global context information at the same time. The memory-efficient attention module and lightweight convolution module were combined to design a new detection head in in order to reduce the number of model parameters and computational cost. The experimental results showed that the mAP@0.5 index of AO-YOLO increased by 2.7% and the mAP@0.5:0.95 index increased by 2.0% on the constructed plateau mountain obstacle dataset (FOD) compared with YOLOv8n, and the overall calculation amount reduced by 24.7%. The proposed model has the characteristics of high accuracy and lightweight in the highland aviation obstacle detection task.