一种基于深度学习的夜间车流量检测方法

当车流量较少时，降低路灯亮度可以达到能源节约目的。为此，采用深度学习中的R-FCN目标检测网络完成夜间车辆检测任务。R-FCN网络相比传统深度学习网络，不仅是基于区域推荐模型的网络，而且引入了平移变化特性，所以对目标检测效果更好。为了占用更少硬件资源，缩小模型规模，采用ShuffleNet通道分组与组间通信机制，压缩原始残差网络。同时，对NMS（非极大值抑制）算法进行修改，从而可以更好地筛选重叠目标，降低网络漏检率。实验结果表明，该方法准确率较高，在UA-DETRAC数据集的夜间图片检测中精度最高可达到90.89%。

Research on Night Traffic Flow Detection Method Based on Deep Learning

When the traffic flow is small， reducing the brightness of the street light can achieve the purpose of energy saving. For night vehicle detection tasks， this paper uses the R-FCN detection network in deep learning. Compared with the traditional deep learning network， the R-FCN network is not only a network based on the regional recommendation model， but also introduces translational variation characteristics， so the network has better effects on detection task. In order to take up less hardware resources and reduce the size of the model， we used ShuffleNet's channel grouping and inter-group communication mechanism to compress the original residual network. At the same time， the NMS algorithm is modified in this paper， so that the overlapping targets can be better filtered and the network's missed detection rate can be reduced. The experimental results show that the accuracy of the method is high， and the accuracy in the night image of the UA-DETRAC data set can reach 90.89%.