Real-time Head Pose Estimation In Driving Situation

Head pose estimation is to extract the rotation and displacement data through processing the input information such as video.For the intelligent driver assistance system,context-aware applications,such as fatigue test,augmented reality based headup display and semantic recognition of the head posture,can be implemented via obtaining the head pose of the driver,which can highly increase the functionality and effect of the driver assistance system.In the complex driving environment,the illumination condition is less desirable and have more changes.Because of the budget and the durability constraints,the on-board computer usually based on mobile platform architecture with relatively inferior computing performance.These factors limit the implementation of the high real-time and high robust head posture estimation algorithm.Among the existing relevant researches,the method based on the appearance model with particle filter and neural network classification is comparatively mature.The method based on the appearance model with particle filter relies on the appearance information and owns the relatively worse robustness when facing the driving environment with complicatedly changing illumination.At the same time,the particle filter engenders more performance burden.Although the classification method of the neural network processes fast,however,the output pose is less precise and discontinuous.Aimed at the problems above,this paper proposed a complete head post estimation system based on the two-dimensional landmarks of human face with correspondent three-dimensional position to solve head pose,thus to get pose estimation result rapidly.The innovation points of the project are concentrated in several following aspects,Firstly,a new algorithm pipeline is proposed to extract face-head information gradually from coarse to fine.Secondly,based on the existing face detection and landmark alignment framework,the pixel intensity comparison feature and local binary feature based methods,which are more adaptable to the driving environment,is used and an algorithm to handle side-view rotation of head is implemented.Last but not least,an instability metric of the pose estimation is proposed and the adaptive recheck method is implemented according to the metric,which trigger the re-detection procedure when needed.In driving situation,it can reduce the numbers of the recheck effectively,which makes the system to have better performance and more stabilized framerate.In the experiment,the systematic algorithm and relative optimization's productivity and practicability were verified through the actual measurement in car and the precision estimation measurement in lab.The results indicated that the head post estimation system proposed in this paper owns strong robustness and outstanding performance thus have good application prospects in driving situation.