Research And Implementation Of Driver Distraction Detection Method Based On Video Surveillance

Drivers' distracted driving is a major safety hazard that causes frequent accidents.Although people have known that distracted driving is very dangerous,the current situation of traffic safety has not been effectively improved because of difficult supervision and the driver's lack of awareness of changing distraction habits.In order to solve this problem,this thesis studies the method of real-time detection of driver distraction using algorithms about image processing and deep learning,based on the driver's surveillance video data.The driver's face and its surroundings are the key areas for the distracted state.In order to eliminate the disturbing background area in the monitoring images and improve the recognition rate of the detection process,this thesis uses a face positioning algorithm to locate the driver's face before the formal detection.Firstly,the algorithm builds the driver's skin color model based on a Gaussian model.Then it uses the Adaboost algorithm to locate the initial face area,and reduces the detection range based on the active area of driver's head,and then segments the driver's skin area based on skin color characteristics.By finding the largest connected region,the noise area is eliminated.Finally,it obtains candidate face regions base on the ares' s center and last face's region,and calculates their confident points.The candidate region with the highest confident point is selected as the new face region.It is improved that compared with the existing KCF tracking algorithm,this algorithm is more stable and accurate according to the experiment.In this thesis,the task of detecting distraction is divided into two phases.In the first phase,only the normal and unnormal driving state are judged.When it is found that there is distracted driving performance,it is further subdivided into categories of its distracted state.In this way,a large amount of detection tasks can be performed in two classifications,which can have higher accuracy than multi-classification tasks.Aiming at this process,a CNN network model for detecting distraction state is proposed.The model is a network that shares some feature extraction parameters and is based on VGG16 and multi-task learning.In addition,by introducing a multi-scale convolution kernel to replace the original VGG single-scale convolution kernel,and using a global average pooling layer instead of a fully connected layer,the calculation speed is furtherimproved.Through experiments,it is proved that the recognition accuracy of the recognition algorithm in this thesis is 89.27% for distinguishing between normal and distracted states,and the accuracy of fine classification for distracted states is 82.63%,and it can also meet the real-time requirements.After researching the algorithm,the thesis finally designs and implements a distraction state detection system.The system can detect the driver's distraction state in real time,and when it is found that the driver is not driving normally,an alarm is issued to the driver,and at the same time,the result information of the detected distraction state category is uploaded to the background database server.