Research On Recognition And Warning Of Driver’s Typical Abnormal Posture

With the increasing number of vehicles and vehicle products,more and more factors influence occurrence of traffic accidents.Traffic accidents due to drivers account for an important proportion.Therefore,studying of the driver’s abnormal behavior in depth,understanding of its internal impact mechanism,and can provide support for reducing traffic safety accidents.The research focuses on driver’s abnormal behaviors such as making phone calls,eating,and losing sight while driving in this thesis.Based on the idea of divide-andconquer,using the vehicle images of the three areas of interest to build models separately.In order to propose abnormal driving behavior recognition and early warning methods,the thesis analyses space as well as time relationships between vertical and horizontal aspects of the abnormal driving behavior.In addition,studying the application of key frame capture method is to improve the real-time performance of the system.The main research work of this thesis includes several aspects:1.Aiming at the performance areas of abnormal behavior on the driver,the research designed an abnormal driving behavior recognition model in three areas of the driver’s eyes,head and upper body.In the specific modeling process,there is a problem of huge video data.The face key points and pose key points can solve the problem of huge calculation amount of video data.This measure can solve the problem of consistency of feature representation and the waste of computing resources caused by repeated application of picture data.In view of the detailed performance characteristics of the three key areas,the implementation scheme combining statistical method and traditional method is adopted.The results show that the recognition accuracy is 84.16%.2.In order to solve the problem of confused early warning,the research establishes a horizontal and vertical classification early warning model after establishing abnormal behavior recognition in this thesis.Based on the speed of completing an action,the importance and interference elimination of triggering early warning are analysed to form a vertical estimate.The horizontal estimate is related to the time and space.This early warning method integrates the importance and relevance of related factors,and improves the practicability of early warning results and the comfort of early warning objects.3.The research proposes a key-frame extraction method based on reinforcement learning in this thesis.In order to reduce the repeated detection of the same information,the paper designed a key frame extraction method.Based on reinforcement learning idea,key frame extraction has the adaptive ability of switching strategy according to the change of driver’s state.The result shows its unique advantages in the existing methods.Finally,the accuracy and real-time of the abnormal driving behavior recognition and early warning model are verified by test analysis,and the results are compared with the driving behavior analysis algorithm of Baidu AI platform.