基于工程环境背景下安全帽佩戴检测算法研究

Research on the Algorithm of Safety Helmet Wearing Detection Based on the Background of Engineering Environment

针对当前安全帽佩戴检测算法存在结构复杂、鲁棒性差等问题,提出一种借助改进后的YOLOv3算法进行安全帽佩戴检测。利用包含通道注意力机制的SE-ResNeXt残差结构,替换YOLOv3模型中Darknet53网络的残差结构,在不加深网络结构的情况下,利用通道注意力机制,捕获特征有用信息,达到提高特征表示能力的目的。再利用空间池化金字塔模块,对待测图片进行多尺度提取,提高检测精度。最后将IOU损失函数替换成CIOU损失函数,在进一步提高检测精度的同时,加速模型收敛。通过自建数据集验证可知,改进后模型检测准确率相比于原始YOLOv3模型,检测平均精确度(mAP)提高了4.29%,每秒检测帧数(FPS)提高了8.67%。

In view of the complex structure and poor robustness of the current helmet wearing detection algorithm, a helmet wearing detection method is proposed with the aid of the improved YOLOv3 algorithm.The SE-ResNeXt residual structure including the channel attention mechanism is used to replace the YOLOv3 model. The residual structure of the Darknet53 network, without deepening the network structure, uses the channel attention mechanism to capture useful information of features to achieve the purpose of improving feature representation. Then use the spatial pooling pyramid module to perform multiscale extraction of the image to be tested. Improved the detection accuracy. Finally, the IOU loss function was replaced with the CIOU loss function, which further improved the detection accuracy while accelerating the model convergence. It can be seen from the self-built data set verification that the improved model detection accuracy rate is compared with the original YOLOv3 model. The average accuracy(mAP) has increased by 4.29%, and the number of frames per second(FPS) has increased by 8.67%.