改进YOLOV5目标检测模型的实时抽烟检测方法

抽烟目标的实时检测难以在实际场景中应用，主要原因是终端设备成本低，能够承载的模型计算量十分有限，难以同时兼顾精度与速度。为此，本文使用改进的YOLOV5目标检测模型，将其双向特征融合网络(FPN+PAN)替换为加权双向特征金字塔网络(BiFPN),增强网络的特征传递能力。实验结果表明，替换特征网络之后的YOLOV5s-BiFPN目标检测模型精度更高，AP_0.5达到91.7%,且模型的参数量、计算量以及FPS基本不变，满足应用在实际场景中进行实时抽烟检测的条件。

Real-time Smoking Detection Method Based on Improved YOLOV5 Target Detection Model

Real-time detection of smoking targets is difficult to be applied in actual scenes, mainly because the cost of terminal equipment is low, and the amount of model calculation that can be carried is very limited, so it is difficult to give consideration to both accuracy and speed. In view of the above problems, this paper uses the improved YOLOV5 target detection model. Its bidirectional feature fusion network (FPN+PAN) is replaced by weighted bidirectional feature pyramid network (BiFPN) to enhance the feature transfer capability of the network. The experimental results show that the YOLOV5s-BiFPN target detection model after replacing the feature network has higher accuracy. The AP0.5 reached 91.7%, and the parameters, calculation and FPS of the model keeps almost constant. Which means that it meets the requirements of real-time smoking detection in actual scenes.