Research On Advanced Driving Assistance System For Driving Recording Video

Since the beginning of the 21 st century,China’s automobile industry has rapidly developed and people have gradually realized their vision of owning cars.However,the increase in the number of cars brings many traffic safety hazards.In order to reduce the incidence of traffic accidents,the Advanced Driving Assistance System(ADAS)has emerged.The emergence of ADAS represents the country’s attention to road traffic safety issues and the future research trends of major automobile companies.This thesis focuses on the research of vehicle detection,lane line detection,lane departure warning,and distance measurement related to ADAS functions.Having a strong data resource is a prerequisite for conducting research,and this thesis takes advantage of the video storage capabilities of a driving recorder to collect a large amount of front-vehicle driving data and lane data in various environments and enhances,expands,and annotates these data.For the experimental part,the research algorithm is integrated using the Jetson Xavier NX development board and installed for operation in the vehicle.The main work of this thesis is as follows:Firstly,in response to the issues of obstruction,difficulty in detecting far-distance targets,and large model size in vehicle detection,an improved YOLOv5-based vehicle detection model is proposed.The backbone network of YOLOv5 is modified to the lighter Ghost Net network.Regarding the defect of NMS missing target boxes in YOLOv5,DIo U-NMS is proposed to solve the target obstruction issue.By replacing the activation function and adding the CA attention mechanism,the model’s precision is improved.Compared with the original YOLOv5,the m AP value of the improved YOLOv5 is increased from 89.28% to 95.50%.Secondly,lane line detection is the premise for conducting lane departure warning experiments.In this thesis,Deeplabv3+ is used as the detection model,and the network backbone is modified to Mobile Netv2 to make the model lighter without losing accuracy.The SE attention mechanism is added to improve model accuracy.Finally,Dense-ASPP is used to replace the original network pyramid structure to enhance feature extraction ability.Through experiments,the improved model’s accuracy reaches 95.8%,and the detection speed reaches55.8.On the basis of lane line detection,the lane departure warning model is proposed by improving the TLC and CCP combined model.According to the technical requirements,lane departure angle and vehicle distance boundary line threshold are set,and the deviation warning function is realized based on thresholds.Finally,based on the improved YOLOv5 vehicle detection algorithm,an improved distance measurement model based on improved pitch angle and yaw angle is proposed.The camera angle is obtained in real-time through the texture road disappearing point algorithm.Then,the pixel coordinates obtained from the improved YOLOv5 are passed to the distance measurement model for coordinate conversion.Experimental results show that the average static road distance measurement accuracy is maintained at 95.54%,and the average moving vehicle distance measurement accuracy is maintained at 93.02%.