基于YOLOv5改进模型的办公室吸烟行为检测

Office Smoking Behavior Detection Based on Improved YOLOv5 Model

为解决目前吸烟行为检测中小目标检测精度低、易误检的问题,提出一种改进YOLOv5的识别模型YO-LOv5s+。该模型将YOLOv5的主干网络与BoTNet相结合,以提高模型的特征提取能力,使其能够检测更小的目标物体;同时改进特征融合部分,在网络模型的颈部应用加权双向特征金字塔BiFPN,以高效融合浅层位置信息与深层高级语义信息,有效提高了检测精度。将网络公开数据集与自制数据集整合成办公室吸烟实验数据集,在该数据集上比较YOLOv5s+模型与原YOLOv5模型的检测性能。实验结果表明,改进模型YOLOv5s+的平均精度均值(mAP)为81.8%,精度为82.8%,召回率为83.9%,相较原模型分别提高了5.4%、4.1%和6.4%,较好地实现了办公室吸烟行为检测。

To solve the problems of low accuracy and easy false detection of small targets in current smoking behavior detection,an improved YOLOv5 recognition model YOLOv5s+is proposed.This model combines the backbone network of YOLOv5 with BoTNet to improve the feature extraction ability of the model,enabling it to detect smaller target objects;At the same time,the feature fusion part is improved by applying a weighted bidirectional feature pyramid BiFPN in the neck of the network model to efficiently fuse shallow position information and deep high-level semantic information,effectively improving detection accuracy.Integrate publicly available online datasets and self-made datasets into an office smoking experimental dataset,and compare the detection performance of the YOLOv5s+model with the original YOLOv5 model on this dataset.The experimental results show that the average accuracy(mAP)of the improved model YOLOv5s+is 81.8%,with an accuracy of 82.8%and a recall rate of 83.9%.Compared with the original model,it has improved by 5.4%,4.1%,and 6.4%,respectively,and has achieved good detection of office smoking behavior.