Privacy Protection-Oriented Research On Driver Head Posture Evaluation Method

With the development of visual artificial intelligence,more and more vehicles are deploying smart cameras to achieve specific functions and services.While smart cameras provide people with a safer and more convenient travel experience,there is also a potential risk of privacy data leakage,although there are many research results about driver status monitoring.However,there are few studies that combine privacy protection and driver status detection methods.Therefore,this thesis combines privacy protection requirements with driver head posture evaluation tasks,and explores the privacy-oriented driver head posture evaluation method.1.In this thesis,we systematically study the effect of mosaicization degree of face region on head pose evaluation on Large Head Poseture Dateset,and innovatively replace the activation function Re LU in the neural network with Swish,so that the network feature mapping map contains more head pose information.Experiments are conducted on data with different degrees of mosaic blurring in the face region,and the results show that the improved algorithm can alleviate the performance degradation of head pose estimation caused by mosaic processing: on data with 75% mosaic degree in the face region,the improved algorithm can improve the head pose estimation accuracy by 6.31° and 3.72° in Yaw and Pitch turning angles cumulative errors,respectively;in Yaw and Pitch The average cumulative error of all experimental groups decreased by about 5° on Yaw and Pitch turning angles.2.Considering the maximum protection of driver’s privacy,the innovative method of topdown driver’s head posture evaluation is proposed: the camera is arranged on top of the driver’s head,and the head posture estimation is performed by acquiring the image of the driver’s head.The top-down camera arrangement can avoid the acquisition of sensitive biometric information.In order to accurately evaluate the head posture from the driver’s overhead image,an ellipse fitting-based image processing algorithm is designed in this thesis.To verify the effectiveness of the algorithm,20 sets of virtual scenes and mannequins are designed jointly with Makehuman and Unity based on the actual driving environment,and the virtual mannequins are set to shake their heads from 60° left to 60° right,and videos are recorded.The verification experimental results show that the estimation error of the top-down head pose estimation algorithm is within 15° for 80% of the data.3.Based on NVIDIA Jetson TX2 platform and ZED camera,the top-down head pose estimation system was built and its performance was verified in both laboratory and real vehicle environments,where the best test results showed that the head pose estimation error was within 15° for 83% of the data.