基于注意力机制改进YOLO-V5的多尺度行人目标检测

全文下载： 20240504.pdf

文章编号： 1672-6987（2024）05-0127-08； DOI： 10.16351/j.1672-6987.2024.05.017

杨旭睿， 冯宇平*， 李悦， 陶康达， 戴家康(青岛科技大学 自动化与电子工程学院， 山东 青岛 266061)

摘要： 为了提高在各类复杂场景中不同尺度行人目标的检测性能，提出了一种结合注意力机制的YOLO-V5多尺度改进算法。通过对YOLO-V5主干网络进行加深，进一步提高其特征提取能力，丰富深层语义信息；在算法中引入Coordinate Attention注意力机制，使其能够关注输入特征图中的有效区域；在原始YOLO-V5基础之上，增加一组新的目标检测头部，来增强算法对小尺度目标的检测性能。所提出的方法在Citypersons行人数据集上进行了实验，将Citypersons验证集中的不同尺度目标细分为3种后，改进算法对这3种不同尺度行人目标的AP50指标分别达到了645%、666%、717%，Recall指标分别达到了530%、566%、617%，较原始YOLO-V5算法分别提高了38%、36%、23%和33%、47%、35%。实验结果表明，提出算法对多尺度行人目标的检测效果具有明显提升。

关键词： 行人目标检测； YOLO-V5； 多尺度目标检测； 注意力机制

中图分类号： TP 391.4文献标志码： A

引用格式： 杨旭睿， 冯宇平， 李悦, 等. 基于注意力机制改进YOLO-V5的多尺度行人目标检测［J］. 青岛科技大学学报(自然科学版), 2024, 45(5): 127-134.

YANG Xurui， FENG Yuping， LI Yue, et al. Multi-scale pedestrian detection based on improved YOLO-V5 combined with attention mechanism［J］. Journal of Qingdao University of Science and Technology（Natural Science Edition), 2024， 45（5）： 127-134.

Multi-Scale Pedestrian Detection Based on Improved YOLO-V5

Combined with Attention Mechanism

YANG Xurui， FENG Yuping， LI Yue, TAO Kangda, DAI Jiakang

(College of Automation and Electronic Engineering, Qingdao University of Science and Technology, Qingdao 266061, China)

Abstract: In order to improve the multi-scale pedestrian detection performance in various scenes, an improved multi-scale YOLO-V5 algorithm combined with attention mechanism is proposed. By deepening the YOLO-V5 backbone network, the feature extraction ability is further improved and deep semantic information is enriched; the Coordinate Attention attention mechanism is introduced into YOLO-V5 to focus on the effective area of the input feature map; a new prediction head is added to the original YOLO-V5 to enhance its detection performance for small targets. The proposed method has been tested on the Citypersons dataset and after subdividing its pedestrian targets in validation set into three different scales, the AP50 values for three different scales pedestrian targets reached 645%, 666%, 717% respectively and the Recall values reached 530%, 566% and 617% respectively, which were 38%, 36%, 23% and 33%, 47%, 35% higher than the original YOLO-V5. The experimental results show that the proposed algorithm can obviously improve the multi-scale pedestrian detection performance.

Key words: pedestrian detection; YOLO-V5; multi-scale target detection; attention mechanism

收稿日期： 2023-11-08

基金项目： 国家自然科学基金项目(61971253);青岛科技大学大学生创新训练计划项目(202410426014).

作者简介： 杨旭睿(1998—),男,硕士研究生.*通信联系人.