基于YOLO v4的夜间车辆检测模型轻量化研究

Research on Lightweight of Night Vehicle Detection Model Based on YOLO v4

针对夜间车辆检测模型的实时性要求,以YOLO v4模型为基础,将主干特征提取网络更改为灵活性强且易于实现的MobileNet V2,并将加强特征提取网络里面的普通卷积全部更改为深度可分离卷积,同时模型给每个通道引入缩放因子,并与该通道输入相乘。然后将缩放因子正则项和权重损失函数联合进行稀疏正则化训练,此时选择较小的缩放因子进行通道剪枝,剪枝后模型的部分通道缺失,检测性能会降低,因此通过模型微调来弥补精度损失,并经过性能评估后再进行修剪迭代。最后得到一个轻量化的车辆检测模型,使其检测速度更快,更能满足夜间车辆检测的实时性需求。经过在UA-DETRAC数据集的实验分析可知:轻量化夜间车辆检测模型的检测精度可达98.29%,同时每秒处理帧数高达42帧图像。

In response to the real-time requirements of the night vehicle detection model,based on the YOLO v4 model,the backbone feature extraction network is changed to MobileNet V2,which is flexible and easy to implement,and changes all the ordinary convolutions in the enhanced feature extraction network to deep separable convolutions.At the same time,the model introduces a scaling factor to each channel and multiplies it with the channel input.Then the scaling factor regular term and the weight loss function are combined for sparse regularization training.At this time,a smaller scaling factor is selected for channel pruning.After pruning,some channels of the model are missing,and the detection performance will be reduced,so the model is fine-tuned to compensate for the loss of accuracy,and after the performance evaluation,the pruning iterations are performed.Finally,a lightweight vehicle detection model is obtained,which makes the detection speed faster and can better meet the real-time requirements of night vehicle detection.The experimental analysis on the UA-DETRAC data set shows that the detection accuracy of the lightweight night vehicle detection model can reach 98.29%,and the number of frames per second can be as high as 42 images.