基于改进YOLOv5s的复杂环境行人检测模型

Pedestrian detection model in complex environment based on improved YOLOv5s

针对行人检测在复杂环境下存在的高误检率和丢失率问题,提出了一种基于YOLOv5s的改进模型YOLOv5s-RFDH。该模型在保留YOLOv5s基线网络的基础上,在特征提取和检测部分进行了优化改进,以提高行人检测在复杂场景中的准确性和鲁棒性。针对CrowdHuman数据集和WiderPerson数据集进行行人目标检测。以上数据集行人密集且存在大量遮挡,因此,采用了K-Means++聚类算法来重新聚类数据集以获取适合数据的锚框;引入感受野模块(Receptive Field Block,RFB)来进行特征提取,在不同分支中使用空洞卷积增加感受野从而提取更深层次的特征信息,并最终将这些特征融合在一起,提升了小目标行人的检测精度;解耦头可以解决目标检测中的尺度不变性问题,引入解耦检测头将分类和回归任务分离,从而能够更加准确地检测到不同尺度和大小的目标。在CrowdHuman数据集和WiderPerson数据集划分出的测试集上进行对比实验,结果表明,改进后的模型在检测准确率上得到提升,丢失率有所下降,在以上两个不同数据集上检测准确率分别提升1.4%和1.2%,丢失率分别降低2.0%和1.7%。

For solving the problems of high false detection rate and missed detection rate of pedestrian detection in a complex environment,an improved model YOLOv5s-RFDH based on Yolov5s is proposed.The model retains the YOLOv5s baseline network while optimizing and improving the feature extraction and detection parts to enhance the accuracy and robustness of pedestrian detection in complex scenes.This paper focuses on pedestrian detection in the CrowdHuman and WiderPerson datasets,which contain dense crowds and substantial occlusions.To address this challenge,the K-Means++clustering algorithm is used to re-cluster the datasets and obtain anchor boxes suitable for the data.The Receptive Field Block(RFB)module is introduced for feature extraction,using dilated convolutions in different branches to increase the receptive field and extract deeper feature information.These features are then fused to improve the detection accuracy of small pedestrian targets.Decoupled heads can solve the scale invariance problem in object detection.By introducing a decoupled detection head that separates the classification and regression tasks,targets of different scales and sizes can be accurately detected.Experimental results on the test sets of CrowdHuman and WiderPerson datasets show that the improved model achieves an increase in detection accuracy and a decrease in missing rate.Specifically,the detection accuracy is improved by 1.4%and 1.2%on the respective datasets,while the missing rate is reduced by 2.0%and 1.7%.These results demonstrate the effectiveness of the proposed algorithm.