多视角图像与PP-YOLOE结合的人群QR码检测方法

A Crowd QR Code Detection Method Combining Multi⁃perspective Images with PP-YOLOE

现有目标检测系统在人群密集场景中无法有效实现尺寸极小快速响应码(QR码)的批量自动化检测,为此,提出一种基于多视角图像与改进PP-YOLOE模型的人群QR码辅助检测方法:首先构建多视角图像采集系统,通过侧视图与顶视图图像完成多种目标归属主体的正确关联;随后在路径聚合网络(PAN)中增加跨层空间注意力模块,提升模型算法小目标检测能力;利用深度可分离卷积对RepResBlock模块进行轻量化改进,提升模型算法执行效率.与其他4种算法的对比实验表明,最优有效目标检测准确率提高9.9%,单次可完成的检测数量达到13个、单目标检测平均耗时72.5 ms.

The existing target detection system is still unable to effectively achieve batch automated detection of extremely small QR(Quick Response)codes in crowded scenarios.To this end,a crowd QR code assisted detection method based on multi per⁃spective images and improved PP-YOLOE model is proposed.Firstly,a multi-perspective image acquisition system is con⁃structed to accurately associate multiple target subjects using side and top-view images.Subsequently,a cross-layer spatial atten⁃tion module is added to the Path Aggregation Network(PAN)to enhance the model's ability to detect small targets.Secondly,the RepResBlock module is lightweight improved using deep separable convolution,which improves the efficiency of the model algorithm execution.In the final comparative experiment,the proposed algorithm outperforms the other four algorithms,achiev⁃ing a 9.9%improvement in effective target detection accuracy.It achieves 13 detections in a single attempt,with an average time of 72.5ms for single target detection.