Study On Driving Behavior Recognition And Risk Assessment Based On Internet Of Vehicle Data

Driving behavior study and risk assessment have important practical significance.Research shows that 75%of accidents are caused by illegal operations,and driving attitude is the decisive factor for traffic safety.If there are some technologies for early warning and regulating driving behavior,80%of accidents could be avoided.17.5%of auto insurance policyholders have frequent traffic accidents and high compensation,which makes 82.5%of the safe driving owners have to pay high premiums.There is still a lot of room for optimization for the car insurance pricing scheme.At present,China is actively promoting the secondary market-oriented reform of commercial vehicle insurance rate,reducing commercial auto insurance rate,further expanding the independent pricing power of insurance companies.Therefore,driving risk assessment will provide technical means for insurance companies to develop personalized auto insurance pricing plans.In this study,smartphone sensor data,time series data provided by the Natural Driving Research Project(NDS),visual data,auto insurance claim data,and other types of IoV(Internet of Vehicle)data is used to conduct driving behavior analysis and risk assessment with deep learning techniques.The main contents are as follows:(1)A novel smartphone sensor data collection and processing approach is introduced,and a deep learning framework for driving dehavior identification and risk prediction model is proposed.The new method eliminates the influence of gravity on smartphone sensor data collection.Subsequently,sensor data is collected and six types of driving events is annotated.To solve the limitation of traditional feature engineering,attention-based DeepConvGRU and DeepConvLSTM was proposed to automatically extract features of driving behaviors.The mechanism behind the method is to capture hidden spatiotemporal characteristics among the high-dimensional sensor data,and automatically extract the key features of driving behavior by fusing Convolutional Neural Network(CNN)and Recurrent Neural Network(RNN).The proposed end-to-end framework was able to learn features from raw data after simple preprocessing and fuse the learned features without any specific feature design.The experimental results indicate that the proposed model outperforms other competing methods significantly.In order to investigate the generality of the DeepConvGRU and DeepConvLSTM,further research and analysis experiments utilizing the NDS time-series data is conducted.By establishing a spectral driving safety risk quantitative standard,relationship between traffic accidents and driving risk levels is established.By constructing the correlation model between crash and driving behavior,the implied spatiotemporal characteristics representing by NDS data is extracted for driving risk level prediction.The experimental results show that the model can suppress the imbalance of dataset effectively,and possesses good generalization ability.(2)A classification and detection method for distracted driving behavior based on visual data is proposed.A framework based on the two-stream 3D convolution network is proposed for distracted driving behavior recognition.Verification experiment compared with the latest optimized 2D convolution network use the public datasets and collected data is conducted.The experiment shows that the two-stream 3D convolutional network can automatically mine spatio-temporal feature and capture temporal context-related information.It is suitable for classification tasks based on time-series image sequences or videos.In order to overcome the impact of background and camera angle on recognition,we re-annotate the classification dataset,produce the detection dataset for distracted driving behavior recognition,that also has general adaptability.We integrate the state-of-the art detection algorithms to achieve the detection of signature behavior,and improve the performance and robustness.(3)An end-to-end deep learning framework named DeepAFM is proposed for driving risk prediction using auto insurance claim data.DeepAFM use Factorization Machine(FM)and Embedding layer to decompose the parameter matrix into two low-dimensional parameter matrices,and the complexity is reduced.It can learn the weight of feature interactions via attention mechanism,which extract such features that are important to the prediction.A fully-connected Deep Neural Network(DNN)was introduced as the deep component of DeepAFM model which can learn high-order feature interactions.