Research On Machine Vision Algorithms For ADAS Perception System

Advanced Driving Assistant System(ADAS)is a low-level stage of the development of driverless vehicles.It mainly includes collision avoidance or pre-avoidance system,lane maintenance system,lane deviation alarm system and so on.Advanced Driving Assistance System can help drivers make correct judgments and play an important role in reducing traffic accidents.Lane detection algorithm and obstacle detection algorithm,as the perception system of ADAS,play an important role in the decision-making of ADAS.The implementation of advanced driving assistance system requires a large number of sensors,including radar sensors,laser sensors,video sensors and so on.Radar sensors and laser sensors are expensive.Video sensors are relatively cheap.Based on machine vision and using video sensors,lane detection algorithm and obstacle detection algorithm for advanced driving assistance system are studied in this paper.The main contents of this paper include:Firstly,the lane line is extracted by using the method of machine vision.By analyzing the different performance of yellow and white lane lines in multiple channels and edge extraction of images,the lane line is extracted as complete as possible by combining various methods,avoiding the missing detection of lane lines,and then the image is penetrated.View transformation into overhead view and quadratic function is used to fit Lane line,which improves the integrity and real-time performance of lane line detection.At the same time,curvature radius of lane line and offset distance of vehicle can be calculated.Secondly,two drivers with different driving abilities are selected to pass the same bend at the same speed,and the video of the driver crossing the bend is collected.The lane detection algorithm is used to process the collected video,and the extracted lane and vehicle driving data are compared and analyzed to verify the accuracy of the lane detection algorithm.Finally,YOLOv3 and SSD based on convolution neural network(CNN)are analyzed and compared.The lengths and widths of large and small objects in YOLOv3 loss function are normalized,and a better YOLOv3-VPC convolution neural network model is obtained.This neural network makes its proportion in the formula become the same,thus improving the ability of network to detect small objects.At the same time,the center loss function is introduced into SSD network,which is combined with the software Max loss function of SSD to improve the robustness of SSD deep learning model.Finally,YOLO-VPC and SSD-VPC network models are obtained,and their mAP values are compared and the reasons for the results are analyzed.