A Study Of Fatigue Detection Methods Based On Video Analysis

It is very significant to study on the detection of human beings espeically for drivers fatigure. A robust and fast human face and eye auto detection method based on Aadboost cascaded classfiers is proposed in this paper, with the image tracking algorithm such as Mean Shift method, it can process real time video of drivers, and detect faces and eyes.and locate eyes to analyse the fatigue status using PERCLOS and times of eye open per minute.It reviewed some faces and eyes detection methods based on features,model matching, and statistical learning.The paper focus on the adaboosting method based on statistical learning method, and fast calculation method of haar features of faces and eyes using integral map proposed by Viola Jones, and how to use cascaded classifiers to construct the fast face and eye detector.Eye detection runs after the completion of face detection,and searches eye targets only in candidate regions after roughly locating based on eye physiological structural features.Since candidate eye region is reduced,the detection will run fast,and more false positive can be reduced,more cascaded layers and stronger classifers can be deployed with less additional CPU time cost.In fatigue detection system,algorithm is required to track eye movement in case of head rotation.light variation,and eye closure.In the paper.Mean Shift method with the Kalman filter combined tracing method was adpoted,the cascaded classifier need not to be applied for each image frame in eye detection. Since searching range is reduced, and redundancy information bewteen frames can be used effectively, the system can run in real time with extensive adaptability.In our experiments, we used one face and two eye cascade classifier libaries provided by OPENCV.The results show good performance in detection rate and execution time. The cascaded classifiers selected in experiment can be upgraded through the training by adding new fresh samples in future applications without changing software frameworks and data structures.It only needs less than two Mega bytes of RAM. Therefore, it fits for applications on embedded system platforms with small memory, and demonstrates potential prospects in future application.