Driver Abnormal Behavior Detection Based On Unsupervised Learning

With the improvement of national living standard,people's demand for self driv-ing travel is increasing day by day,and the automobile industry has been developing rapidly.But the development of automobile industry,while facilitating people's travel,also leads to more and more traffic accidents.The incidence of traffic accidents caused by drinking water,smoking and other abnormal driving behaviors has increased year by year.Therefore,it can effectively ensure the driver's travel safety by detecting the driver's abnormal driving behavior accurately and quickly and giving early warning to the driver.In this paper,according to the driver behavior habits,the task of abnormal driving detection is divided into the abnormal detection of the mouth area and the eye area of the driver.Then,according to the feature of two different areas,different ab-normal detection models are designed to detect abnormal driving behavior.Finally,an experimental platform of driver abnormal driving detection system based on embedded system is built to verify the practical application value of the model.The main research work and innovation points of this paper are as follows:1.Because of the complex abnormal situation in the mouth area,the traditional anomaly detection method can't detect abnormality Precisely.Therefor,this paper pro-poses a driver's mouth anomaly detection algorithm based on multi-scale convolutional AutoEncoder network.Firstly,the algorithm uses the facekey points detection network to obtain the area of driver's mouth quickly,removing the irrelevant background.Then,the improved CAE algorithm is used to reconstruct the picture of the mouth area and calculate the reconstruction error to determine whether there is any abnormality.The following three improvements are included:adding skip-connect structure to better re-tain shallow features;The Inception structure is added and the proportion of branch channel is adjusted to fit the input image better;The robust of model detection is im-proved by adding Gaussian white noise in training process.The experimental results show that compared with the traditional AutoEncoder algorithm,the AUC is improved from 0.682 to 0.938,and can run on the embedded system in real-time.2.There are two main features in eye region images:low resolution and abnormal situation mainly occurs in the center area of the region.According to those features,this paper proposes a new feature descriptor named GW-CCH.Comparing to traditional Contrast Context Histogram Featurethe,GW-CCH uses Gaussian weighted algorithm to calculate the average gray value of the eye region,which increases expression of the eye region];Then,FindCBLOF,a local anomaly factor detection algorithm based on clustering,was used to detect the abnormalities occured in eye region.The experimental results show that GW-CCH is more suitable for the expression of the eye region than other handcraft features and Findcblof algorithm is the best method in the abnormal detection of eye region.3.In order to test the above algorithms,this paper builds a real-time detection system for driver abnormal driving based on embedded platform to verify the practical application value of the algorithm.