基于改进YOLOv8的复杂环境行人识别模型

Pedestrian Recognition Model for Complex Environment Based on Improved YOLOv8

摘要: 针对现有行人识别模型在复杂场景下，存在小目标及密集目标误检漏检率高等问题，提出了基于YOLOv8改进的行人检测模型Swin-YOLO。首先设计了SE-Conv模块，该模块通过将卷积操作和注意力机制进行融合，有效提升了模型的特征选择能力；其次，引入Swin Transformer (ST)模块，利用ST模块大感受野特性提升模型的长距离建模能力；最后，在输出端构建了基于深度卷积与空洞卷积的多尺度卷积模块，该模块旨在以较低的计算复杂度实现高效的多尺度特征融合。实验结果表明，在小目标场景数据集TinyPerson中，Swin-YOLO模型相比YOLOv8模型在检测精度指标mAP@0.5和mAP@0.5:0.95上分别提升了6.5和2.5个百分点，有效减少了误检和漏检，为解决智能辅助驾驶及道路状况预警提供一种有效的改进YOLOv8行人识别模型。

Abstract: To address the high false positive and false negative rates of existing pedestrian detection models in complex scenes, particularly for small and densely clustered targets, we propose Swin-YOLO, an improved pedestrian detection model based on YOLOv8. First, the SE-Conv module was designed, which effectively enhances the model’s feature selection capability by integrating convolutional operations with attention mechanisms. Second, the Swin Transformer (ST) module was introduced to leverage its large receptive field for improved long-range modeling. Finally, a multi-scale convolutional module based on deep convolutions and dilated convolutions was constructed at the output layer to achieve efficient multi-scale feature fusion with reduced computational complexity. Experimental results demonstrate that on the TinyPerson dataset for small-object scenarios. The Swin-YOLO model outperforms the YOLOv8 model by 6.5 and 2.5 percentage points in mAP@0.5 and mAP@0.5:0.95 respectively. This effectively reduces false positives and false negatives, to address intelligent driving assistance and road condition warnings, an effective improved YOLOv8 pedestrian detection model is provided.