针对行人目标检测任务中，存在遮挡、尺度变化和背景复杂等问题带来的检测性能降低的情况。基于YOLOv7-tiny网络，设计了一种行人目标检测网络DR-YOLO（deformable region YOLO）。首先，在ELAN（efficient layer aggregation network）模块中融入轻量化可变形卷积DCNv4（deformable convolution network v4）来建立长短范围内的特征信息的交互，以增强模型对行人特征的表达能力。其次，在颈部使用一种多尺度特征融合结构，用于融合不同尺度目标的特征，增强模型对尺度变化的适应性，并降低参数量，同时还在网络主干加入改进的注意力机制，提高模型对感兴趣目标的关注度来优化后续网络的特征提取。从而提高模型在面对复杂情况时的检测性能。实验结果表明：本文方法在CrowdHuman和WiderPerson构成的数据集上mAP_0.5和mAP_0.5-0.95分别取得了80%和46.8%的检测效果，与原模型相比分别提高了1.8%和2.7%，参数量减少17.2%，与现阶段同类算法相比，在没有参数量增加的情况下，提高了模型的检测精度，展现出一定的优越性。

In the task of pedestrian target detection, challenges such as occlusion, changes in scale, and complex backgrounds can significantly degrade detection performance. Based on YOLOv7-tiny network, an improved pedestrian target detection algorithm DR-YOLO (deformable region YOLO) was proposed. First, the ELAN (efficient layer aggregation network) module was integrated with the lightweight deformable convolution DCNv4 (deformable convolution network v4) to establish interaction between long-range and short-range feature information, thereby enhancing the model's ability to express pedestrian features. Second, a multi-scale feature fusion structure was used in the neck to fuse the features of different scale targets, enhancing the model's adaptability to scale changes and reducing the number of parameters. In addition, an improved attention mechanism was added to the main body of the network to enhance the model's focus on interesting targets, thereby optimizing the subsequent network's feature extraction. This improved the model's detection performance in complex situations. Experimental outcomes indicate that the methodology introduced herein attains mAP_0.5 scores of 80% and 46.8% for mAP_0.5-0.95, when evaluated on a dataset comprising CrowdHuman and WiderPerson. Relative to the baseline model, these findings denote enhancements of 1.8% and 2.7% respectively, concurrently with a diminution of 17.2% in parameter count. Compared with similar algorithms at the current stage, without increasing the number of parameters, the detection accuracy of the model has been improved, demonstrating certain superiority.