Study On Driver Unsafe Behavior Detection Based On RGB-D

With the continuous increase of highway mileage and car ownership,it not only promotes the development of transportation,but also causes frequent traffic accidents.However,90% of traffic accidents are caused by driver behavior,including unsafe operations such as phone calls and playing mobile phones.Therefore,the detection of drivers' unsafe behavior has become a hot topic in the current intelligent traffic system.Most of the current research methods are based on RGB data for detection,which is susceptible to interference such as light changes and background occlusion,and has many defects in practical applications.With the development of depth cameras,the cost of RGB-D cameras is reduced and the accuracy is improved.More and more researchers have proposed to use RGB-D cameras to study behavior detection algorithms.This paper mainly studies the detection method of driver unsafe behavior based on RGB-D.Driver behavior detection consists of data collection and preprocessing,feature extraction and feature fusion classification.Taking these three factors into consideration,in order to extract robust features more efficiently,this paper proposes a RGB-D-based driver unsafe behavior detection method.The main work is as follows:First,for the data collection and preprocessing,This paper uses RGB-D calibration data for registration to improve data reliability.Then,for the hole noise of depth image due to interference,this paper uses a RGB image-based depth image hole repair algorithm,using the similarity of color information to filter neighboring depth values to fill holes,comparing morphology and filtering methods can be effective on depth images hole repair and denoising.After that,in order to reduce the background interference in feature detection,this paper proposes the RGB-D + Vi Be driver region of interest extraction algorithm,results show that the accuracy of this method is more than 90% under different lighting environments,and the robustness and accuracy of the algorithm are significantly higher than the Vi Be algorithm based on a single image.Secondly,three different feature modules of arm feature,head feature and eye feature are designed as driver behavior feature in the feature extraction link.The coordinate and depth value of the arm key bone point in RGB-D image is selected as the output in the arm feature extraction module.,And six facial key points are selected to establish a three-dimensional head model to solve the head pose Euler angle as the head feature module output,then the three methods of circular Hough transform,circular Gabor filtering and brightness feature matching are used to accurately locate the iris of the driver's eye area,and according to iris position and the eye corner,the eye open-close state and the estimated line of sight is judged as output of eye feature module.Thirdly,for the feature classification,this paper uses Real Ada Boost strong learning algorithm to design the classifier.In order to improve the robustness of the classifier,this paper chooses the decision tree as the weak classifier to reduce sensitivity to missing values.In this paper,the method based on rgb-d multi-feature +Real Ada Boost is tested on 250 test samples of simulated cockpit collected in the laboratory,The experimental results show that the driver's unsafe behavior detection accuracy rate is more than 90%,and the driver behavior type detection accuracy rate is more than 85%,which can effectively detect the driver state.The classifier experimental results show that the Real Ada Boost classifier performs better than other algorithms.In the feature comparison experiment,the accuracy of the multi-feature classification algorithm based on RGB-D is significantly higher than that of the multi-feature algorithm based on RGB or single feature.Meanwhile,The arm feature based on RGB-D is the feature that most significantly affected the detection of drivers' unsafe behaviors among all the features in this paper.In this paper,we study the detection of drivers' unsafe behaviors based on RGB-D.The integrated color and depth information improves the accuracy and robustness of the detection,and our method can be applied to driving safety assistant to warn drivers of unsafe behavior and guarantee road safety effectively.