Abstract:

The identification of abnormal driving behavior plays a vital role in traffic safety, and accurate identification of abnormal driving behavior can significantly improve driving safety. At present, the detection and identification of abnormal driving behaviors, such as rapid acceleration, rapid deceleration, sudden left turn or right turn, etc. In the process of vehicle driving mainly adopts video surveillance or clustering method. In these two methods, the actual effect of the former is restricted by the application scene, while the latter cannot identify the driving behavior of specific individual vehicles. In order to solve the above problems, an extended neural network detection model based on Bi-LSTM and FC is proposed. The model can effectively use the characteristics of sudden change of driving data in time series and improve the recognition accuracy of abnormal driving. The neural network model can effectively use the time series characteristics of driving data and accurately identify the abnormal driving behavior of vehicles. The accuracy rate can reach 98.08%.

Key words: deep learning, driving behavior recognition, Bi-directional Long Short-Term Memory（Bi-LSTM）

摘要：

异常驾驶行为的识别对交通安全起着至关重要的作用，准确识别异常驾驶行为能够显著提高驾驶安全。目前，针对车辆行驶过程中的异常驾驶行为，如急加速、急减速、突然左转或右转等的检测识别，主要采用视频监控或聚类的方法完成。在这两种方法中，前者的实际效果受到应用场景的制约，而后者则不能针对具体的单个车辆进行驾驶行为识别。针对以上问题，使用一种基于双向长短记忆网络（Bi-LSTM）及全连接神经网络（FC）的拓展神经网络检测模型，该模型能有效利用行车数据在时间序列上发生突变时的特征，提高异常驾驶行识别准确率。将车辆行车数据处理后制作数据集并对模型进行训练，训练完成后的神经网络模型能够有效利用行车数据的时间序列特征，准确识别车辆的异常驾驶行为，准确率可达到98.08%。

关键词: 深度学习, 驾驶行为识别, 双向长短记忆网络