人员安全帽佩戴轻量化检测方法研究

Research on the lightweight detection method of person helmet wearing

为有效减小安全帽检测算法的计算复杂度,并提高算法对于小目标的检测精度,提出一种基于Pytorch深度学习框架的轻量化安全帽检测模型。使用轻量化网络设计减小模型的计算量;设计可变形双向聚合网络提高模型对检测对象尺度和形状多样性的适应能力,优化对小目标的检测效果;通过网络收集的施工现场图像验证安全帽检测算法的检测效果。与已有安全帽检测算法相比,该方法检测精度有明显提高、模型参数量显著下降,并以137帧/s的速度运行。可变形双向聚合网络利用深层语义特征和浅层细节特征,并自适应调整感受野,可以适应不同形状和尺寸的检测对象,提高检测精度。

The existing helmet detection algorithms based on deep learning have high computation complexity and high requirements for hardware computing capabilities, resulting in a high cost in actual applications. Moreover, the diversity of the shape and scale of the detection targets are not fully considered in existing algorithms, so there also exist common problems of missed and false detection of small targets. Aiming at the above problems, this paper proposes a lightweight helmet detection model based on the Pytorch deep learning framework to reduce the amount of calculation of the model. In addition, a deformable bi-direction aggregation network is proposed to transfer shallow detail information and deep semantic information in a bi-directional way, thereby improving the model’s adaptability to detect targets of different scales. And the deformable convolution is introduced to improve the model’s adaptability to detect targets of different shapes. We validate the effectiveness of our algorithms with extensive experiments on a helmet detection dataset(Safety Helmet Wearing Dataset, SHWD). We use 6 000 images in SHWD as the training set and 2 000 images as the testing set. A computer with Intel-Xeon(R) 4214 CPU(2.2 GHz), 64 GB memory, and four NVIDIA GeForce GTX2080Ti GPUs are used as the experimental platform. The experimental results indicate that the recognition accuracy of the proposed algorithm is over 90%. The proposed deformable bi-direction aggregation network employs deep semantic features and shallow detail features, and adaptively adjusts the receptive field, which can adapt to objects of different shapes and scales, thus improving the detection accuracy. Besides, the proposed algorithm achieves a detecting speed of 137 frames per second, which far meets the needs of real-time operation. Compared with the existing helmet detection algorithms, the detection accuracy and running speed are significantly improved, and the parameters are significantly reduced.