The invention discloses a method and a system for identifying violation behaviors based on an optimized yolov4 algorithm. The method comprises the following steps: (1) carrying out model training through the optimized yolov4 algorithm, and extracting an optimal model; the optimization of the yolov4 algorithm comprises the following steps: reducing the number of channels of an input feature map of a previous layer to one fourth of the original channel number by adopting 1 * 1 convolution in an input part; generating output features of receptive fields of different sizes by using four cavity convolution branches of different sizes; fusing the output feature information of the four different branches by using a receptive field fusion structure; adopting 1 * 1 convolution in the convolution layer of the output part to adjust the channel number of the output feature map; and obtaining a final output feature map by using a residual network; (2) collecting video image data as the input of the optimal model, and detecting violation behaviors; and (3) generating real-time early warning information according to a detection result, and returning the real-time early warning information. According to the invention, the target identification precision in the violation identification image can be improved, and a real-time early warning function is realized.
本发明公开了一种基于优化yolov4算法的违章行为识别方法及系统,包括以下步骤:(1)通过优化yolov4算法进行模型训练,抽选出最优模型;yolov4算法的优化包括:在输入部分采用1×1卷积将前一层的输入特征图的通道数减少到原来的四分之一;使用四个不同大小的空洞卷积分支来生成不同大小感受野的输出特征;利用感受野融合结构融合四个不同分支的输出特征信息;在输出部分的卷积层采用1×1卷积来调整输出特征图的通道数;利用残差网络获得最终的输出特征图;(2)采集视频图像数据作为最优模型的输入,进行违章行为的检测;(3)根据检测结果生成实时预警信息,将实时预警信息回传。本发明能够提升了违章识别图像中目标识别精度,并且实现实时预警功能。
Method and system for identifying violation behaviors based on optimized yolov4 algorithm
一种基于优化yolov4算法的违章行为识别方法及系统
2022-01-25
Patent
Elektronische Ressource
Chinesisch
Pedestrian Detection Algorithm Based on Improved YOLOv4
IEEE | 2023
|UAV Target Detection Algorithm Based on Improved YOLOv4
Springer Verlag | 2022
|Method and device for identifying vehicle violation behaviors based on V2P
Europäisches Patentamt | 2023
|