To address the limitation of single-view detection in complex underground environments, a novel dual-branch network architecture for air-ground cooperative detection is proposed in this paper. The proposed method consists of two independent encoder-decoder branches, which are used to extract aerial and ground features separately. A cross-attention mechanism is introduced in the encoders to enable feature interaction between different viewpoints. Furthermore, a dynamic weight fusion module is designed to adaptively fuse aerial and ground features, and a joint loss is incorporated for backpropagation to optimize both the branches and the fusion module. Experimental results show that the dual-branch DETR achieves a 19.8% improvement in mAP compared to dual-branch Faster R-CNN. Ablation and inference experiments further validate the effectiveness of the proposed method. The results demonstrate that the Transformer-based framework is more suitable for improvements in dual-branch cooperative detection.