基于图像序列分析的城市道路交通事故预测  

Urban Road Traffic Accidents Prediction Based on Image Sequence Analysis

作　　者：胡正华 周继彪 毛新华[4] 张敏捷[1] HU Zhenghua;ZHOU Jibiao;MAO Xinhua;ZHANG Minjie(School of Cyber Science and Engineering,Ningbo University of Technology,Ningbo 315211,Zhejiang,China;Department of Transportation Engineering,Tongji University,Shanghai 201804,China;Ningbo Highway Construction&Management Center,Ningbo 315199,Zhejiang,China;College of Transportation Engineering,Chang'an University,Xi'an 710064,China)

机构地区：[1]宁波工程学院,网络空间安全学院,浙江宁波315211 [2]同济大学,交通运输工程学院,上海201804 [3]宁波市高等级公路建设管理中心,浙江宁波315199 [4]长安大学,运输工程学院,西安710064

出　　处：《交通运输系统工程与信息》2024年第5期91-102,共12页Journal of Transportation Systems Engineering and Information Technology

基　　金：浙江省哲学社会科学规划课题成果“交通事故黑点治理的数字化探索”(22NDQN279YB);宁波市自然科学基金(2023J185)。

摘　　要：为进一步提高路网交通事故预测的准确性,本文提出一种基于图像序列分析的短时交通事故预测方法。首先,使用过采样方法对微信小程序采集的交通事故数据进行插值处理,来消除事故数据内部大范围零值对模型训练准确性的影响;并将其与路网的流量数据及与引发事故相关的属性数据进行融合,得到稳定的时间序列作为模型的输入数据。然后,构建一个具有密集连接卷积的双向ConvLSTM U-Net(BCDU-Net)模型。该模型结合了一组双向ConvLSTM结构,将编码层和解码层的特征进行融合,以更全面地捕捉序列数据中的时空相关性。同时,模型还采用密集连接卷积结构,使特征图在深度方向上相互拼接,确保每一层都能够直接从损失函数中访问梯度。最后,通过将预测结果与实际交通事故数据的比较评价了模型的性能。结果表明,本文模型的预测结果相比全连接长短期记忆网络(FC-LSTM)模型,卷积长短期记忆网络(ConvLSTM)模型和U-Net模型,交叉熵损失函数分别降低了65.96%、15.70%和3.47%,均方根误差分别降低了21.48%、3.13%和1.71%,精确度分别增加了75.06%、11.82%和3.08%。说明本文所提出的方法在预测城市道路交通事故时具有更好的性能。To further improve the accuracy of traffic accident prediction in road networks,a short-term traffic accident prediction method based on sequential image analysis is proposed.First,an oversampling technique is applied to interpolate traffic accident data collected from a WeChat mini-program to mitigate the impact of extensive zero values within the data on model training accuracy.These data are then integrated with road network traffic flow and accident related attributes to generate stable time series as input for the model.A Bidirectional ConvLSTM U-Net with densely connected convolutions(BCDU-Net)is constructed.In this model,bidirectional ConvLSTM structures are used to integrate the features from the encoder and decoder layers,comprehensively capturing spatiotemporal correlations in the sequential data.Additionally,densely connected convolutions are employed to concatenate feature maps in the depth dimension,ensuring that each layer can directly access gradients from the loss function.Finally,the performance of the proposed model is evaluated by comparing the predicted results with actual traffic accident data.The results show that,compared to the Fully Connected Long Short-Term Memory(FC-LSTM)model,the Convolutional LSTM(ConvLSTM)model,and the U-Net model,the proposed model achieves reductions in cross-entropy loss of 65.96%,15.70%,and 3.47%,reductions in root mean square error of 21.48%,3.13%,and 1.71%,and increases in precision of 75.06%,11.82%,and 3.08%,respectively.It is demonstrated that the proposed method offers superior performance in predicting urban road traffic accidents.

关 键 词：智能交通 交通事故预测 深度学习 交通安全 BCDU-Net模型 

分 类 号：U491.1[交通运输工程—交通运输规划与管理]