卷积神经网络及其在智能交通系统中的应用综述

从特征传输方式、空间维度、特征维度3个角度，论述了近年来卷积神经网络结构的改进方向，介绍了卷积层、池化层、激活函数、优化算法的工作原理，从基于值、等级、概率和转换域四大类总结了近年来池化方法的发展，给出了部分具有代表性的激活函数对比、梯度下降算法及其改进型和自适应优化算法的工作原理和特点；梳理了卷积神经网络在车牌识别、车型识别、交通标志识别、短时交通流预测等智能交通领域中的应用和国内外研究现状，并将卷积神经网络算法与支持向量机、差分整合移动平均回归模型、卡尔曼滤波、误差反向传播神经网络、长短时记忆网络算法从优势、劣势和在智能交通领域的主要应用场景三方面进行了对比；分析了卷积神经网络在智能交通领域面临的鲁棒性不佳和实时性较差等问题，并从算法优化、并行计算层面和有监督学习到无监督学习方向研判了卷积神经网络的发展趋势。研究结果表明:卷积神经网络在视觉领域具有较强优势，在智能交通系统中主要应用于交通标志、车牌、车型识别、交通事件检测、交通状态预测；相比其他算法，卷积神经网络所提取的特征更加全面，有效地提高了识别准确度与速度，具有较大的应用价值；卷积神经网络未来将通过网络结构的优化、算法的改进、算力的提升以及基准数据集的增强，为智能交通带来新的突破。 

Review of convolutional neural network and its application in intelligent transportation system

From the perspectives of feature transmission mode, spatial dimension and feature dimension, the improvement directions of convolution neural network structure in recent years were reviewed. The working principles of the convolution layer, pooling layer, activation function and optimization algorithm were introduced, and the recent developments of pooling methods in terms of value, level, probability, and transformation domain were summarized. The comparison of some representative activation functions, and the working principle and characteristics of the gradient descent algorithm and its improved and adaptive optimization algorithm were given. The application and research status of convolutional neural network in intelligent transportation fields such as license plate recognition, vehicle type recognition, traffic sign recognition, and short-term traffic flow prediction were reviewed. The convolutional neural network algorithm was compared with the support vector machine, differential integrated moving average regression model, Kalman filter, error back propagation neural network, and long-term and short-term memory network algorithms from the advantages and disadvantages and main application scenarios in the field of intelligent transportation. The issues of poor robustness and poor real-time performance of convolutional neural network in the field of intelligent transportation were analyzed. The development trend of convolutional neural network was evaluated in terms of algorithm optimization, parallel computing, and supervised learning to unsupervised learning. Research results show that the convolutional neural network has strong advantages in the field of vision. It is mainly used for traffic sign, license plate, vehicle type recognition, traffic event detection, and traffic state prediction in intelligent transportation system. Compared with other algorithms, the convolutional neural network can extract more comprehensive features. It can effectively improve the recognition accuracy and speed and has great application value. The convolutional neural network will bring new breakthroughs to intelligent transportation in the future through the optimization of network structure, the improvement of algorithm and computing power, and the enhancement of benchmark data sets. 5 tabs, 3 figs, 146 refs.