Research And Implementation Of Image-sensitive Target Detection Technology For Unmanned Combat Vehicles

Object detection is one of the key tasks in the field of computer vision.This subject takes the images captured by unmanned ground vehicles in the field as the research object,explores how to detect each field target,to estimate which category it belongs to,and to locate its boundary box.Finally we design and implement a complete object detection system.The specific contents are as follows:(1)First,we make an acquisition scheme of field object data set.The field object data set used in this paper is made mainly through two ways,camera acquisition and web spider.We analyze the characteristics of the field objects,and find that the field objects in the pictures collected by the unmanned ground vehicles are often confusing,occluded,and remote.In order to deal with these characteristics of objects and detect field objects with higher speed and higher accuracy,we design a detection scheme in this paper.(2)Second,we propose and implement a data enhancement solution based on deep convolutional Generative Adversarial Network,DCGAN.In our method,residual module is introduced into DCGAN,so we can deepen the discriminator and improve the network's ability to extract the features of images.Thanks to this,the discriminator can distinguish between the true and false samples more strictly,which in turn force the generator better learn the real distribution and generate images more similar to the real samples.The experimental results show that the method can effectively enhance the characteristics of the field object data set in view of the shortcomings of limited collection scenarios,insufficient quantity and lack of content.Besides,we comprehensively use other data enhancement methods to enrich the content of the field object data set and finally solve the problem of difficult collection of training data.(3)Third,we design and implement an object detection model based on YOLOv3.In this paper,we select the model of YOLOv3 since it shows the best performance.We also make a series of improvements to it and make the developed model more suitable for our subject.These improvements mainly include: K-means++ clustering algorithm is used to generate the prior box suitable for the field object data set;In the original YOLOv3,MSE was used as the loss function of boundary box regression,while in consideration of the deficiency of MSE in evaluating the prediction of boundary box,the loss function of GIo U is used instead in this paper,so that the network has higher detection accuracy;According to the requirements of realtime and low missing rate in the field object detection tasks,this paper adjusts the penalty item of the confidence loss function so that the network can achieve a higher recall rate.The effectiveness of these improvements have been verified by experimental results.(4)Fourth,we employ transfer learning method to optimize the training process of the model.Specifically,we first use large public data sets,COCO and VOC to train the backbone of the whole detection network,Darknet-53.Next,we use the data set in this paper with the size of 224 x 224 to fine-tune the network.Then,we adjust network to 448 x 448,using the data set to continue training Darknet-53 classification network.Finally,we use the trained weights in Darknet-53 to initialize the front part of the model,and use random initialization method to initialize the other parameters,and train detection network as a whole.Experiments show that multi-scale training can improve the performance of the network,and multi-scale detection can improve the detection accuracy of the network.(5)Finally,a field object detection system based on the research content of this paper is introduced.The hardware requirements and design points of the field object detection system are analyzed,and the main functional modules are introduced.