摘要
为了有效改善传统交通事故预测方法过分依赖于交通管理者主观经验的不足,结合计算机视觉领域的深度学习框架,提出了一种基于ConvLSTM模型的交通事故预测方法。首先将城市路网划分为规则格网,然后统计每个格网中累计发生的交通事故数并作为一个图像的像素值,将由此形成的时间序列数据对构建的模型进行训练并用于短时交通事故的预测。该模型可以有效地捕捉城市路网交通事故在时间和空间上的分布特性,克服了传统FC-LSTM网络模型忽略交通事故空间分布特性的缺陷。通过对比实验证明:ConvLSTM网络模型具有较好的预测准确率和精确度,在用于预测短时交通事故方面具有更好的应用前景。
In order to effectively improve the deficiencies of traditional traffic accident prediction method which relies too much on the subjective experience of managers,a traffic accident prediction approach based on ConvLSTM is proposed in combination with the deep learning framework in the field of computer vision.This method firstly divides the road network into the urban areas into regular grids,and then counts the cumulative number of traffic accidents in each grid as a pixel value of an image.The resulting time series data are used to train the ConvLSTM model and predict short-term traffic accidents.ConvLSTM can effectively capture the spatial and temporal characteristics of traffic accidents in the road network,and overcome the shortcomings of traditional FC-LSTM,which ignores the spatial characteristics of accidents.It is proved through comparative experiments that the ConvLSTM model has better prediction accuracy and precision.Therefore,ConvLSTM has a better application prospect in predicting short time traffic crashes.
作者
郭旭
胡正华
GUO Xu;HU Zhenghua(School of Cyber Science and Engineering,Ningbo University of Technology,Ningbo 315211,China)
出处
《宁波工程学院学报》
2023年第1期9-15,共7页
Journal of Ningbo University of Technology
基金
浙江省哲学社会科学规划课题(22NDQN279YB)。