| 摘 要: 针对现有的高精度行人检测模型因资源要求高而导致的难以应用于边缘计算场景的问题,提出了一种 适用于边缘GPU设备的轻量级实时密集行人检测算法。该算法通过在检测头中融合全维度动态卷积,降低了冗余 信息对于检测效果的影响,并通过优化损失函数增强了算法区分待检测目标和背景的能力。实验结果表明,在密集 人群场景下的行人检测任务中,该算法在精确度方面较本文基线算法YOLOv7-tiny提升了4.1百分点,这证明该算 法能够在边缘计算场景下实现准确的密集人群检测。 |
| 关键词: 行人检测;小目标识别;深度学习;边缘计算 |
|
中图分类号: TP391
文献标识码: A
|
|
| Lightweight Pedestrian Detection Algorithm for Dense Crowd Scenes based on YOLOv7-tiny |
|
ZHANG Xinyuan1, GAO Zhigang2, FENG Jianwen1
|
(1.School of Computer, Hangzhou Dianzi University, Hangzhou 310018, China;
2.College of Inf ormation Engineering, China Jiliang University, Hangzhou 310018, China)
zhxy@hdu.edu.cn; gaozhigang@cjlu.edu.cn; fengjianwen@hdu.edu.cn
|
| Abstract: In response to the challenges posed by existing high-precision pedestrian detection models, which require substantial resources and are thus difficult to apply in edge computing scenarios, this paper proposes a lightweight real-time dense pedestrian detection algorithm suitable for edge GPU devices. This algorithm reduces the impact of redundant information on detection performance by integrating full-dimensional dynamic convolution in the detection head, and enhances the algorithm's ability to distinguish between the target to be detected and the background through optimizing the loss function. Experimental results demonstrate that in pedestrian detection tasks within densely populated scenes, this algorithm improves accuracy by 4.1 percentage points compared to the baseline algorithm YOLOv7-tiny presented in this paper, proving that it can achieve accurate dense crowd detection in edge computing scenarios. |
| Keywords: pedestrian detection; small object recognition; deep learning; edge computing |
