Driver Cell Phone Usage Detection Based On Semi-supervised Support Vector Machine

With the development of the times,unmanned driving has become the hottest topic today,but the only essential step in truly unmanned driving is the widespread use of Advanced Driver Assistance Systems(ADAS).The ADAS system mainly uses a plurality of sensors to collect information,and carries out risk assessment on the collected information through target detection,identification and tracking and other processing technologies,so as to determine in advance possible dangers for the driver and ensure the safety of driving.In the construction of the ADAS system,I focused my research on driver cell phone usage detection.During the development process,we trained a classifier by collecting a large number of samples to judge driver behavior.However,a large number of sample collections consumes a lot of manpower and material resources.In particular,the collection and labeling of driver cell phone usage detection has brought a lot of tedious work.Faced with this difficulty,this article attempts to solve this problem through semi-supervised learning using a small number of mark samples.The main work of the dissertation is as follows:(1)For the problem of driver cell phone usage detection area division,it is proposed to determine the call detection area according to the deflection of the head position,so that the obtained rectangular box area information will be more valuable,thus ensuring the excellentness of the final detection result..(2)When the current semi-supervised support vector machine detects the call behavior,the data caused by the call sample is far less than the normal driving sample,resulting in the problem that a large number of call samples are mistakenly divided into normal driving samples.In this paper,by improving the local search algorithm,the boundary values that may be misclassified during classification are processed twice.Finally,a semi-supervised support vector machine based on driver cell phone usage detection is obtained.Finally,the effectiveness and robustness of the method are proved by experimental comparison.(3)It is unclear whether the ratio of the actual calling sample to the normal driving sample for the driver cell phone usage detection video leads to the problem that the class ratio of the mark sample is uncertain.In order to ensure the good stability of the marker samples under different class ratios,the authors analyzed the data and found that the data conforms to the distribution of the DBSCAN clustering algorithm.Using this feature,a semi-supervised DBSCAN was proposed and used to estimate the unlabeled samples.The true positive and negative sample proportions,semi-supervised support vector machines obtained by using the sample proportion information were integrated with the semi-supervised DBSCAN to obtain the final results.Experiments show that the algorithm has the advantages of high performance and strong stability.