基于时间卷积自编码网络的实时交通事件自动检测方法

为了提高道路异常交通事件检测效率并降低误报率,提出了一种基于时间卷积自编码网络的实时交通事件自动检测方法。首先设计了基于波动相似性度量的交通模式搜索算法用来筛选具有相同交通规律的样本数据;并构造了交通流模式矩阵作为网络模型输入,以避免样本不均衡与单一样本数据随机性对交通模式学习的干扰;同时设计了新的时间卷积自编码网络对交通模式特征进行无监督提取并对未来交通参数进行合理预测;为了降低交通流参数随机波动性带来的事件判别的干扰,设计了异常状态评估方法,通过对模型预测误差分布的学习,结合当前检测数据给出最终的事件判定结果。采用美国西雅图I90公路与I405公路2015年全年的交通流检测数据与历史事故数据进行实证研究,并与6种典型交通事件检测算法进行性能对比。研究结果表明:基于时间卷积自编码网络的实时交通事件自动检测算法具有较高的检测率、较低的误报率以及更快的平均检测时间;综合各种交通运行情况下,可接受误检率分别为5%、10%时,平均检测率可分别达到93%、98%;同时算法能够自适应学习交通状态的动态变化,对不同交通运行环境具有较强适应性与稳定性。

Real-time Automatic Method of Detecting Traffic Incidents Based on Temporal Convolutional Autoencoder Network

A real-time traffic-incident automatic detection method based on a temporal convolutional autoencoder network was developed to improve the detection efficiency of abnormal traffic incidents and reduce the false alarm rate. First, a traffic pattern search algorithm based on the volatility similarity technique was designed to screen the sample with the same traffic pattern, and the traffic flow pattern matrix was constructed as the network model input to prevent the interference of sample imbalance and single-sample data randomness on traffic pattern learning. A new temporal convolutional autoencoder network was designed to perform an unsupervised extraction of traffic pattern features and reasonable predictions of future traffic parameters. An abnormal-state evaluation method was developed to minimize the interference caused by the random volatility of traffic flow, and combined with the current detection data, could yield the final incident discrimination results through modeling prediction error distribution. The traffic flow detection data and historical accident data of Seattle I-90 and I-405 highways obtained in 2015 were used for empirical research, and the proposed method was compared with several typical traffic incident detection algorithms. The comparison results show that the real-time traffic-incident automatic detection using temporal convolutional autoencoder has a comparative higher detection rate, lower false alarm rate, and a faster average detection time. Under various traffic operation conditions, the average detection rate can reach 93% and 98% when the acceptable false detection rates are 5% and 10%, respectively. This method can adaptively learn the dynamic changes in the traffic state and exhibits strong adaptability and stability to different traffic-operating environments.