Research And Implementation Of Driver Unsafe Behavior Recognition Model Based On Deep Learning

With the development of modern transportation,more and more motor vehicles are driving on the roads,and people are becoming more and more convenient.At the same time,it has also brought a series of road safety issues.These safety problems mainly focus on the driver's negligence and overconfidence,and the inducement is often the driver's unsafe driving behavior.The Ministry of Transport and the Sichuan Provincial Department of Transportation issued more detailed road video surveillance specifications in 2018 and 2019,respectively.The Ministry of Transport's specification emphasizes that road video surveillance systems should include monitoring of driver driving behavior.The specification of the Department of Transportation in Sichuan Province is more focused on the specific test indicators of terminal equipment and communication protocols in terms of active safety.It is the general trend that provinces across the country issue detailed specifications in accordance with local conditions.This thesis refers to the two authoritative specifications that have been issued.From the perspective of active safety,it studies the automatic identification and warning of driver's unsafe behaviors to regulate driver behaviors,and reduces their travel risks.This thesis designs a recognition system in a practical application scenario.Its system implementation consists of an algorithm analysis module,a result feedback module,and a data management module.The algorithm analysis module is responsible for automatic identification and alarming.The result feedback module is used for automatic identification.The results are manually corrected and fed back,and used for iterative incremental learning of the algorithm analysis module to promote product updates.The data management module is used to manage all data and generate statistical tables.In the end,a demo version of the system was initially implemented to show the related functions.This thesis confirms the structure of the data set of the research object of Wuliangye surveillance video data according to engineering requirements,and divides the data into training set,validation set,test set A and test set B.The training set,the validation set and the test set A follow the same sample distribution and belong to data set A.They use the content-based key frame extraction algorithm proposed in this paper,which is suitable for training data.The test set B uses another method of random sampling of keyframes,the result of which is closer to the actual data,for data set B.The pre-processing module preprocesses the video to generate a single frame or a sequence of frames,and the target detection subsystem extracts the positioning information of the face area,related object areas,and special areas from the image data.The classification and recognition subsystem performs feature fusion on this information,and uses a single frame model(real-time version)or a time series model(non-real-time version)to output the results.In the end,among the three behaviors of calling,smoking,and not wearing a seat belt,the real-time version reached 99.56%,99.13%,98.77%(test set A)and 66.74%,63.88%,63.85 on data set A and data set B,respectively.The non-real-time version reached 99.81%,99.55%,98.45%(test set A)and 70.06%,67.43%,70.15%(test set)on dataset A and dataset B,respectively.The real-time version also achieved a test accuracy of 95.96% on the Egypt data test set,which is equivalent to 96.31% of the best model in the industry,indicating that the design has certain generalization capabilities.