基于BH-YOLOX算法的电动车头盔检测系统研究

Research on Electric Vehicle Helmet Detection System Based on BH-YOLOX Algorithm

摘要: 为了提升电动车头盔检测系统的精确度和实时性，考虑到传统检测方法在特征提取方面存在困难，目前基于深度学习的目标检测算法存在模型的泛化性不高，对复杂背景的适应性不足等问题。本文提出基于YOLOX算法优化后的BH-YOLOX模型。首先，特征提取网络维度，通过构建轻量级的Ghostnet网络，在减少了参数量和计算量的前提下，同时提高了特征提取能力，使模型更加轻量化；特征融合维度，增加了Squeeze-and-Excitation Networks (SENet)通道注意力机制，加强了不同通道的特征的关联，提高了网络在复杂场景中的性能；扩展特征层维度，BH-YOLOX在原有的三个特征层的基础上又增加了一个更大的特征层，能够有效提升网络对小目标的检测性能；最后优化损失函数，提高网络模型的回归精度。实验结果证实，BH-YOLOX模型的mAPz值达到98.90%，检测速度为104.51 FPS，能满足绝大多数交通场景的要求，也适用于部署在如摄像头等边缘设备上。

Abstract: In order to improve the accuracy and real-time performance of the helmet detection system for electric vehicles, considering the difficulties in feature extraction, the current target detection algorithms based on deep learning have some problems, such as low generalization of the model and insufficient adaptability to complex background In this paper, the optimized BH-YOLOX model based on YOLOX algorithm is proposed. First of all, the feature extraction network dimension, through the construction of a lightweight Ghostnet network, in the premise of reducing the number of parameters and calculations, while improving the feature extraction capability, makes the model more lightweight. For the feature fusion dimension, the channel attention mechanism of Squeeze-and-Excitation Networks (SENet) is added, which strengthens the correlation between the features of different channels and improves the performance of the network in complex scenarios. By expanding the dimension of the feature layer, BH-YOLOX adds a larger feature layer on the basis of the original three feature layers, which can effectively improve the detection performance of the network on small targets. Finally, the loss function is optimized to improve the regression accuracy of the network model. The experimental results confirm that the mAPz value of the BH-YOLOX model reaches 98.90% and the detection speed is 104.51 FPS, which can meet the requirements of most traffic scenes and is also suitable for deployment on edge devices such as cameras.