Research On Unmanned Target Detection Method In Complex Weather Based On Deep Learning

The driverless technology can greatly improve the traffic efficiency,improve the utilization rate of vehicles and reduce the rate of traffic accidents,thereby greatly improving the level of social productivity,which plays an important role in promoting the development of society and the improvement of people’s quality of life.Unmanned target detection under complex weather conditions is an important part of current unmanned driving technology.In order to make the unmanned driving technology more widely and intelligently applied to people’s production and life,how to improve the detection accuracy of unmanned targets under complex weather conditions is the top priority of research.At present,the detection of unmanned targets under complex weather conditions mainly involves image distortion and blurring,which leads to the problem that unmanned targets are missed or mistakenly detected.In order to solve these problems,traditional methods use dynamic vision,multi-sensor fusion and other methods to improve the detection accuracy of targets under complex weather conditions.Based on the above problems and research,this paper proposes an unmanned target detection algorithm in complex weather based on attention mechanism.The main research contents are:1.To solve the problem that there is no public data set of unmanned driving targets in complex weather on the current network,this paper first collects images and labels them to obtain the original data set.For the problems of unbalanced categories and scenes of the original dataset,this paper uses traditional geometric transformation and pixel transformation methods,DCGAN depth confrontation generation network method and improved UNET method to enhance and balance the data of the original dataset,and finally obtains an excellent dataset DCW2022 with rich images,diverse scenes and balanced categories.2.Unmanned target detection in complex weather conditions is one of the most challenging tasks in the field of target detection.It requires accurate detection of unmanned targets in rain,snow,fog,sand and dust.In order to improve the target detection accuracy,this paper uses the mixed domain attention mechanism CBAM module to improve the network’s ability to analyze the image edge contour,thus improving the ability to extract image features;The two-stage serial fusion attention mechanism group SASNet module is designed to further enhance the network’s ability to extract features of targets and improve the network’s ability to detect small and weak targets,so as to improve the network’s detection accuracy for unmanned targets in complex weather conditions.3.Based on the strategy in 2,this paper proposes a multi-level tandem hybrid domain target detection network MHNet based on YOLOv5 network.This paper first uses the common dataset VOC2007 and MS-COCO to verify the performance of MHNet and conduct visualization experiments.Then use the self-built dataset DCW2022 to verify the performance of MHNet and conduct visualization experiments.Finally,five kinds of complex weather scene images in the DCW2022 data set are tested and visualized.Finally,the experiment proves that MHNet has a good detection effect for unmanned targets in complex weather.