Recognition And Classification Of Dangerous Driving Behavior Based On Machine Vision

With the construction of a large number of traffic roads and the increase in traffic demand produced by people’s daily life,the number of cars and drivers is also increasing year by year.However,the consequent complication of road traffic and difficulties in standardizing driver’s driving behavior are also increasing.In recent years,even if the transportation department has actively taken measures to supervise and regulate the driving behavior of drivers,traffic accidents still occur from time to time due to the objective reasons of the cost of supervision and the subjective reasons of the driver’s lack of driving literacy.At present,vehicles that can design and install driver assistance systems often cost more,which makes it difficult for driver assistance systems to be widely used.In addition,the existing quantitative methods of distracted driving behavior have different measurement standards and data collection methods,which makes it difficult for their risk quantification models to be recognized and widely applied.Therefore,the topic of this article is designed to design methods and applications for the recognition,supervision and rating of driving behavior as the research goal.The main work done is summarized in the following four aspects:1.This thesis proposes an image data enhancement method for the cab scene.In the driving scene,the cab has problems with rapid changes in lighting and reduced data quality caused by body shaking.So the method consists of two parts.One is to design a color enhancement method that limits the contrast to balance the global illumination for the uneven illumination when the illumination is sufficient and the image is too dark when the illumination is insufficient.The second is to design a single-scale fuzzy kernel estimation method for the vehicle environment based on the existing image deblurring method.And the experiment proves that simplifying the fuzzy kernel estimation process can reduce the deblurring processing time of the equipment.2.This thesis uses the superior recognition ability of the VGG network model and the ResNet network model in the field of image recognition to design a dual-network joint behavior recognition method based on deep learning.Because the effect of the fusion of multiple network models in the deep learning method is better than that of a single network model.Therefore,this thesis uses the model fusion method to combine the results of the two base models to improve the effect of behavior recognition.At the same time,in order to reduce the degree of confusion between classes when the joint model is recognized,this thesis adds a spatial self-attention mechanism to the base model training to increase the weight of the key areas of the data.Experiments show that the joint model proposed in this thesis is better than a single model on the driving behavior data set.3.This thesis proposes a risky driving attenuation scoring model based on driving behavior.By analyzing the problems of the existing dangerous driving assessment model in practical application,a quantification model of the degree of danger attenuation based on driving behavior recognition combined with the influence of time and speed is designed.The attenuation scoring model has high fault tolerance and adaptability,so it has a certain reference value in practical applications.4.Design and implement a prototype system for dangerous driving detection and classification.The image data enhancement method,dual-network joint recognition method and attenuation scoring model proposed in this thesis are combined with real-time application scenarios to develop a prototype system.The realization of the prototype system proves that the methods and models proposed in this article have certain application value.