摘要
针对目标检测算法对小目标行人识别率低、对监控远处视野目标检测精度不理想的问题,提出了改进YOLOv5高效多尺度特征利用的行人检测算法。首先,通过在原网络中改进高效的特征融合结构,提高模型对深层特征的感知力来提高模型精度;其次,采用Res2Net Block重构骨干网络,加强对细粒度特征信息的利用;最后,加入改进的空间金字塔注意力池化网络,强化模型的多层次特征表达能力。在CrowdHuman数据集进行训练和验证,YOLOv5-SA的平均检测精度达到了85.6%,相比原算法提高了3.8%,检测速度可以达到51 FPS(frames per second),识别精度和检测速度均具有较好的效果,可以有效应用于密集目标行人检测任务。
Aiming at the problems that the target detection algorithm has low recognition rate of small target pedestrians and unsatisfactory detection accuracy for monitoring distant field of view targets,in this paper we propose a pedestrian detection algorithm with improved YOLOv5 efficient multi-scale feature utilization.Firstly,by improving the efficient feature fusion structure in the original network,the perceptiveness of the model for deep features improves the model accuracy;secondly,the Res2Net Block is used to reconstruct the backbone network to enhance the utilization of fine-grained feature information;finally,an improved spatial pyramid attention pooling network is added to strengthen the multi-level feature representation capability of the model.Trained and validated on the CrowdHuman dataset,the average detection accuracy of YOLOv5-SA reaches 85.6%,which is 3.8%higher than the original algorithm,and the detection speed can reach 51 FPS,with good results in both recognition accuracy and detection speed,it can be effectively used for dense target pedestrian detection tasks.
作者
吴迪
宋家豪
李睿智
WU Di;SONG Jiahao;LI Ruizhi(College of Physical Science and Technology,Shenyang Normal University,Shenyang 110034,China)
出处
《沈阳师范大学学报（自然科学版）》
CAS
2023年第6期536-541,共6页
Journal of Shenyang Normal University:Natural Science Edition
基金
国家自然科学基金资助项目(11804235)。
关键词
小目标行人
注意力模块
密集行人检测
空间金字塔池化网络
small target pedestrian
attention module
occlusion pedestrian detection
spatial pyramid pooling network
