Taxi Driver Violation Detection Based On Deep Learning

With the increasing demand for people’s travel,taxis play a vital role in people’s travel because of their convenience,comfort and speed.However,using mobile phones and smoking during driving threaten the safety of passengers seriously.These behaviors also affect the ride experience of passengers.At present,for the detection of violations,manual screening is the most widely used method.But due to fatigue or distraction,this method can’t guarantee the continuity and reliability of screening and it also consumes a lot of manpower.This paper uses image processing technology to realize real-time automatic detection of taxi drivers’ violations,which is of great significance to the public and supervisory authority.Based on the detailed analysis and comparison of current deep learning algorithms on target detection,this paper constructs a real-time automatic detection method for taxi drivers’ violations based on SSD target detection algorithm.Firstly,we collect samples of drivers’ violations in taxi scenes and label them according to the format of VOC2007 data set after selection and pretreatment.The targets of data set include two categories: mobile phones and cigarettes,totaling 10 981 pictures.After labeling,we make statistics of the scale and aspect ratio of the target in the data set to lay a foundation for the establishing and training.Then,we chose VGG16 as the basic network of SSD algorithm.The parameters of the prediction layer are set according to the small-scale change of the self-built data set and the unique aspect ratio.At the same time,in order to solve the problem of easy-to-divide negative samples in the data set,the loss function is replaced by focal loss.After the modification,we use the method of transfer learning to train data sets.The accuracy of the model can reach 94.22%,which is increased by 2.42% compared with the original SSD algorithm.The processing speed of the model can reach 33 frames per second.Finally,in order to improve the processing speed of the model,SqueezeNet with simpler structure is selected as the basic network of SSD algorithm.The processing speed of the model can reach 55 frames per second on the premise of guaranteeing the accuracy of 89%.Rainbow-SSD algorithm improves the accuracy to95.37% by increasing the number of output feature graphs in the prediction layers.RefineDet algorithm improves the accuracy to 96.16% by combining the idea of two-stage and one-stagemethods and increasing the filtering operation of a priori box.Compared with the original algorithm,the accuracy of the two algorithms is improved by 1.15% and 1.94%,respectively.The results of this paper have been applied in Xining taxi management system and achieved good results.The application shows that the results can realize the automatic detection of taxi drivers’ violations,which can save a lot of manpower,improve the efficiency of the supervision department and provide a guarantee for citizens’ travel.