Abstract:

In recent years, with the rapid growth of car ownership in China, the traffic pressure and safety hazards caused by cars are increasing year by year, so it is particularly important to supervise the behavior of vehicles on the road. Therefore, an improved Yolo_v2 convolutional neural network is proposed for vehicle detection. Firstly, the structure of Yolo_v2 network is improved, residual network is added to improve the accuracy, and multi-scale layer is added to improve the detection accuracy of different size targets in the image. Secondly, Kelu activation function is designed based on the ELU activation function to further improve the detection accuracy. Thirdly, multi-directional vehicle data set and license plate data set are produced. Finally, the vehicle detection system and vehicle are combined. The license plate detection system is integrated into ROS system and communicates with QT creator visual interface, so as to observe the experimental results more clearly. The test results show that the improved Yolo_v2 convolution neural network has a superior performance in traffic violation monitoring.

Key words: Yolo_v2, residual network, multiscale, activation function, multi-directional, violation monitoring

摘要：

近年来，我国随着汽车保有量增长迅速，由汽车造成的交通压力及安全隐患也逐年递增，对道路上车辆行为进行监督也变得尤为重要。提出了利用改进的Yolo_v2卷积神经网络来进行车辆检测。对原Yolo_v2网络进行结构改进，添加残差网络来提高检测准确率，添加多尺度层来提升对图片中不同尺寸目标的检测精度；基于Elu激活函数设计出Kelu激活函数，进一步提高检测准确率；制作多方位车辆数据集及车牌数据集；将车辆检测系统与车牌检测系统集成到ROS系统中，并与QT-Creator可视化界面通信，以便更清晰地观测实验结果。实验结果表明，利用改进后的Yolo_v2卷积神经网络对道路上的车辆进行违章检测有着优越的表现。

关键词: Yolo_v2, 残差网络, 多尺度, 激活函数, 多方位, 违章检测