基于改进YOLOv5和视频图像的车型识别

Vehicle Type Recognition Based on Improved YOLOv5 and Video Images

为了提高车型识别的精度和检测速度,提出了改进YOLOv5的车型识别算法。首先利用高速公路收费的监控视频数据扩充BIT-Vehicle车型数据集,同时针对数据集中各车型图片数量不均衡现象利用图像翻转、添加高斯噪声、色彩变化等图像处理技术对各车型数量进行均衡化,构建BIT-Vehicle-Extend数据集;其次,添加RFB(receptive field block)模块用于增加网络感受野,有助于模型捕捉全局特征;第三,将无参数的SimAM注意力机制添加Bottleneck中,在不增加参数的情况下,提高网络的特征提取能力。实验结果表明,相比于原始网络模型,本文所提出的YOLOv5优化算法,mAP0.5和mAP0.5:0.95达到98.7%和96.3%,分别提高了0.7%和1.5%。在检测速度方面,达到90 frames/s,与原网络相比检测速度基本不变。因此,本文所提出的YOLOv5优化算法,能够高精度的实时检测车型信息,满足车型识别检测需要。

In order to improve the accuracy and detection speed of vehicle type recognition, an improved vehicle recognition algorithm of YOLOv5 was proposed. Firstly, the BIT-Vehicle dataset was expanded by using the monitoring video data of highway toll collection. At the same time, aiming at the imbalance of the number of pictures of each vehicle in the data set, the number of each vehicle was balanced by using image processing technologies such as image flipping, adding Gaussian noise and color gamut change, and the BIT-Vehicle-Extend dataset was constructed. Secondly, the RFB module was added to increase the network receptive field, which was helpful to capture the global features of the model. Thirdly, the parameter free SimAM attention mechanism was added to bottleneck to improve the feature extraction ability of the network without adding parameters. The experimental results show that compared with the original network model, the proposed YOLOv5 optimization algorithm, mAP0.5 and mAP0 5: 0.95 to 98.7% and 96.3%, increase by 0.7% and 1.5% respectively. In terms of detection speed, it reaches 90 frames/s, which is basically unchanged compared with the original network. Therefore, the YOLOv5 optimization algorithm proposed in this paper can detect the vehicle information in real time with high precision and meet the requirements of vehicle identification and detection.