Research And Implementation Of Driving Behavior Recognition In Intelligent Internet Of Vehicles

Autonomous driving is one of the development trends of intelligent transportation.Information transmission,storage and computation in autonomous driving technology rely on Internet of Vehicles(IoV).Until fully autonomous driving(ie level-4),Advanced Driver Assistant System(ADAS)still exist.ADAS can monitor the driving behavior of the driver in real time.When distracted driving behavior occurs,the driver is reminded in real time,or automatic driving takes over manual driving,which can effectively avoid traffic accidents.There are still many problems in driving behavior recognition that can’t be effectively solved,such as high recognition accuracy,high real-time performance,and edge computing based on the Internet of Vehicles(IoV).This thesis studies the driving behavior recognition based on Internet of Vehicles(IoV).The deep convolutional neural network(CNN)is adapted to extract useful information of driver’s visual data and output the classification of driving behavior,such as: safe driving,drinking,using mobile phone,etc.The research work of this thesis is as follows:1.Collection and processing of driving behavior data.Due to the lack of public datasets,and the environmental factor of light is not considered,we have established our own dataset.In the process of making the dataset,we took into account the effects of light,driver,vehicle operation and props(such as mobile phones,drinking glasses,etc.),and expanded the dataset.An image brightness detection algorithm was implemented,and the effect of demotion blur and histogram equalization preprocessing on data quality was verified.2.A two-way attention based driving behavior recognition method is proposed,which is suitable for vehicle image data.This method simulates the human visual system,focusing on image areas that are strongly related to driving behavior recognition,and ignoring image areas that are not related to driving behavior recognition as much as possible.With this end-to-end trainable two-way attention module,the recognition ability of distracted driving behavior is improved.3.Contrastive experiments were conducted on the three situations of no attention mechanism model,one-way attention mechanism and two-way attention mechanism,and the effectiveness of the model was further verified by comparison with other driving behavior recognition algorithms.Through the experiment of the influence of light changes in the driving environment,the adaptability of the model to light changes and the effectiveness of the preprocessing algorithm used in this paper are verified.The experimental results show that the adoption of the two-path attention mechanism has a certain improvement in recognition effect compared to the no-attention mechanism model and the one-way attention mechanism.Both histogram equalization and blind de-motion blur preprocessing algorithms are effective in improving the model recognition effect.The experiment proves the effectiveness of the dual attention mechanism model combined with the preprocessing algorithm.4.Realization of driving behavior recognition based on Internet of Vehicles(IoV).Two implementation methods of driving behavior recognition based on Internet of Vehicles(IoV)are discussed: driving behavior recognition based on web,and driving behavior recognition based on offline mobile device.Using the Flask and PyTorch Mobile framework,the implementation and testing of the server-side system and offline applications are completed.