Research On Optimization Method Of Driver Violation Detection In Taxi Scene

With the rapid development of online car-hailing,it has brought a lot of convenience to people’s lives,but traffic accidents caused by drivers’ violations also occur frequently.The automated real-time monitoring of taxi drivers’ violations is of great significance to the operation and management of taxis and the construction of civilized cities,which can effectively reduce the economic losses caused by traffic accidents.The current detection of taxi drivers behavior is mainly to monitor whether the driver has violated regulations.Traditional research methods include sensor-based research,but sensor-based research suffers from too much passenger interference information,resulting in low test accuracy.In the research based on machine vision,many researches on violations have not reached the desired results in accuracy and speed.In this paper,the target detection algorithm is used to detect illegal targets around the driver,mainly to detect two kinds of illegal targets,cigarette butts and mobile phones,and to jointly track the trajectory of the detected targets on the result of target detection to judge whether the driver has illegal behavior.And optimize the model of the misdetected and missed images in the target detection,and finally realize the illegal behavior detection of taxi drivers based on target detection and target trajectory tracking.The work of this article is mainly reflected in the following aspects:1.Establish a dataset of taxi drivers’ violations.Through the investigation of violations in traffic scenes,this article focuses on the detection of violations of taxi drivers.It is found that taxi drivers have various violations during driving,and the definition of violations of drivers in taxis is carried out.The relevant national regulations and policies were interpreted in detail.Selected to study two important violations of taxi drivers,smoking and using mobile phones,and based on this,we established the research basis of this article-taxi driver violation detection data set.2.Using the YOLOv4 model for target detection,and optimizing the model is a better model detection effect.Using transfer learning to compare the performance of the current mainstream target detection algorithms in this data set,the YOLOv4 target detection algorithm was finally selected.Aiming at the problem that the YOLOv4 algorithm misdetects the steering wheel as a mobile phone in the taxi drivers’ violation detection test set,the Anchor box and loss function are improved,which improves the target detection accuracy by2.13%.In addition,the detection speed of the model in the detection of taxi driver violations has been improved.Combined with the characteristics of the YOLOv4-Tiny network and the lightweight YOLOv4 network structure,the detection speed of the model has been improved to a certain extent.The model can detect 226 more photos per second on the data set in this article.3.In the result of target detection,track the detected target.After target detection,this article found that some images in the data set contained targets,but the driver did not have corresponding violations.When a single target detection is used,this situation will be detected as a driver’s violation.On this basis,this paper tracks the trajectory of the target,and at the same time detects whether the driver’s hands and head interact with the target object to determine whether the driver has violated the rules.The addition of target trajectory tracking increases the accuracy of driver violation detection without losing detection speed,and finally makes the progress of taxi violation detection effectively increased to 97.35%.