Abstract: In view of the situation that vehicles are partially blocked by green belts, roadblocks and other vehicles and encounter low lighting, which brings great difficulty to the field image detection of vehicles, a convolutional neural network algorithm was proposed to detect vehicles. The geometric model of the vehicle was defined based on the computer vision system, including the shape of the vehicle, the symmetry and the shadows it produced. Target detection and data processing technology of convolutional neural network algorithm in deep learning was adopted as the algorithm of the proposed system. The convolutional neural network was applied to detect obstacles in front of vehicles. By comparing different sampling convolutional neural networks, a convolutional neural network to detect vehicles ahead was designed. Through the design and testing of the proposed system, the algorithm of detecting the vehicle in front can ensure the rapidity and the accuracy. The experimental results show that the convolutional neural network with overlapping sampling and pooling layer of Max+Ave+Ave structure has relatively high accuracy and stability.

Key words: traffic engineering; convolutional neural network; picture processing; vehicle detection

摘要： 针对车辆局部被绿化带、路障、其它车辆遮挡以及低照明等问题，给现场车辆图像检测带来困难，提出了基于卷积神经网络算法对车辆进行检测。基于计算机视觉系统构建了车辆的几何模型，包括车辆的形状，对称性以及和车辆阴影。采用深度学习中卷积神经网络算法的目标检测和数据处理技术，将卷积神经网络应用于检测车辆前方障碍物，对不同采样的卷积神经网络结构进行比较，设计了能够检测前方车辆的卷积神经网络算法。并通过系统试验测试，检测前方车辆的算法能够保证检测快速性和准确性。试验结果表明：采用重叠采样和池化层为Max+Ave+Ave结构的卷积神经网络具有较高的车辆检测准确率和较强的系统稳定性。

关键词: 交通工程；卷积神经网络；图像处理；车辆检测