Research On Intelligent Wiper Control Method Based On Machine Vision And Convolutional Neural Network

In rainy and snowy weather,in order to ensure a clear view in front of the driver,all types of vehicles need to be equipped with wiper devices.Due to the fact that traditional wipers need to be manually turned on,they can distract the driver’s attention and cause problems such as delayed response,which can affect driving safety.With the continuous development of computer and image processing technology,using image processing technology to detect rainfall using vehicle mounted cameras has become a new idea.In this paper,machine vision is introduced into cars to sense rain conditions,and vehicle mounted cameras are used to collect images in front of car windshields to determine rainfall.Realizing that automobile wipers can perform corresponding wiper operation modes according to different rainfall conditions is a technological innovation that conforms to the proposed concept of advanced driving assistance system ADAS.Not only can it give drivers a better driving experience,but it can also free them from complex driving responsibilities.In view of this,two types of raindrop detection schemes are proposed,which are implemented based on target detection and classification detection.Based on real-time rain detection for deep learning target detection,a more general rain detection scheme is proposed using rain recognition and rainfall determination methods based on deep learning classification detection.Firstly,a real-time rain detection system and wiper control method based on deep learning target detection are proposed.Through deep learning,the detection and quantity statistics of raindrops are realized on an embedded platform.Using the detected rainfall information,the images captured by vehicle mounted cameras are divided into several categories: no rain,light rain,moderate rain,and heavy rain.The detection results can be used to control the speed of the wiper motor,and the amount of rain corresponds to the wiper operating modes of different gears to achieve an automatic wiper control system.A training model was established using a CNN structure,using YOLOv5 for raindrop target detection.Due to the large differences between daytime and nighttime images,in order to improve recognition performance,separate training was conducted for daytime and nighttime images based on the principle of data diversity filtering,and an all-weather model suitable for daytime and nighttime was further designed.At the same time,a model optimization strategy was adopted to optimize the loss functions GIOU,DIOU,and CIOU,improve the Mosaic data enhancement algorithm,and introduce the Adam optimizer.In order to verify the performance of the system,tests were conducted on the embedded platform Jetson TX2.The accuracy rate and recall rate of the all-weather model reached a high level of 0.86 and 0.83,and the processing speed on the embedded system can reach 20 frames per second,confirming the practicality of the proposed system.Secondly,by analyzing the characteristics of rain that covers the entire windshield in extreme weather,we further design a rain classification detection method that can be applied to any rain form in extreme weather.Analysis shows that in extreme weather,rainfall patterns vary and are in an irregular state.Design rain classification and detection models for Res Net,VGG-16,and Mobile Net,while combining transfer learning to improve strategies and optimize attention mechanisms.The classification model effectively solves the problems of missed and false detection of raindrops in previous(based on raindrop counting)raindrop detection work,while eliminating the interference of complex background on raindrop detection.In view of the lack of extracting and analyzing rainfall information from images of complex rain conditions in extreme weather in target detection,a raindrop image recognition algorithm based on an improved Res Net depth residual network is proposed to achieve a rain detection scheme that can adapt to any rain pattern,achieving a detection accuracy of 95%.Finally,an experiment was conducted on a real vehicle to verify the effectiveness of the classification model,and the experimental results were compared,analyzed,and validated.