Improvement Of Facial Landmarks Detection Algorithm Along With Research Of Real-time Drowsiness Detection Algorithm

With the rapid development of society,traffic safety issue has emerged,among which,unsafe driving conductions,such as fatigue driving,has caused a large number of traffic accidents and severely threatened people's lives and property.Thus,applications of modern science and technology to drowsiness detection own great significance.Among many technologies,computer vision-based technology has attracted wide attention from academia and industry due to its simple equipment and high popularity.In particular,facial landmarks detection technology offers the possibility of real-time and robust drowsiness detection.This thesis focuses on the improvement of facial landmarks detection algorithm based on cascade regression tree and the drowsiness detection algorithm based on facial landmarks.Firstly,the Newton boosting tree model is extended to general multi-response learning problems.Through cascaded multi-response Newton boosting tree model,a facial landmarks detection model is established.Meanwhile,this thesis also proposes constructing shape indexed pixel statistics feature and the corresponding feature selection method,so as to describe the distribution of the pixels in the local areas of the image.In order to meet the real-time requirement,the facial landmarks are triangulated and a Delaunary descriptor is constructed.Then a subset of Delaunary descriptor is selected based on mutual information.In the light of that,this thesis establishes a unified classification model for yawning and sleeping and proposes a unified fatigue detection algorithm.The experiments show that multi-response Newton boosting tree along with the statistics feature helps reduce the error of face feature point detection without sacrificing too much time efficiency.Meanwhile,the proposed fatigue detection algorithm based on facial landmarks has good real-time performance and robustness and can meet the needs of practical application scenarios.