Research On Driver Dangerous Driving Behavior Detection And Early Warning System

With the gradual improvement of China’s traffic infrastructure,it not only brings convenience to people,but also brings a lot of road traffic injuries.The biggest cause of death caused by traffic accidents is the driver’s fault.If the driver has improper driving behavior for accurate identification and timely warning,the probability of traffic accidents can be reduced to a great extent.Compared with the single image recognition method,continuous behavior recognition can achieve higher accuracy,so this paper takes driving behavior as the research object,and uses the method based on video key frame extraction and behavior recognition to study driving behavior recognition.Firstly,because there are few public data sets of existing driving behavior,and most of the public data sets have usage restrictions,which are not enough to support model training,this paper constructs the DB-dataset driving behavior dataset.According to NHTSA’s definition of distracted driving,combined with the actual situation in China and referring to the categories of driving behavior in the State Farm dataset,a new driving behavior dataset is redesigned and collected,which provides data support for the subsequent driving behavior recognition model.Secondly,to solve the problem of low recognition efficiency caused by redundant frames in driving behavior video,an adaptive key frame extraction method is proposed in this paper.Based on the deep learning method based on motion features,Mobile Net-V2 is used to extract driving behavior feature diagrams.To enhance the adaptability of Mobile Net-V2 network to scale,the Inception module in Google Net is improved,and the common Bottleneck structure is replaced by multi-scale feature fusion structure to make the features extracted by the model richer in spatial scale to obtain more detailed feature map output,and then output key frames based on the peak value of MSE curve.After the key frame sequence of driving behavior is obtained,the recognition model of driving behavior is built.In order to solve the problem that the video information can not be fully expressed only by image features in driving behavior video,a hybrid attention mechanism driving behavior recognition method based on fusion of key frames is proposed in this paper.Firstly,a lightweight double-flow convolution model is built by using Mobile Net-V2 network.In order to increase the recognition accuracy,a mixed attention mechanism is introduced,which is added to the beginning of each Bottleneck block of Mobile Net-V2 to average all its channels to obtain global spatio-temporal features.Then the output of the Softmax layer is fused to build a hybrid attention mechanism driving behavior recognition model with key frames.Finally,a set of portable driving behavior recognition and early warning system is designed by combining the algorithm with software and hardware.The embedded vehicle system integrates four functions: driving video acquisition,driving behavior identification,dangerous driving behavior early warning and recognition result analysis.The background management system of driving behavior includes various functions such as driver information management,driving behavior data analysis and so on.In practical application,the driving behavior recognition system designed in this paper provides a guarantee for the safe driving of drivers.