Image Processing And Transmission In Cooperated UAV Search And Rescue System

The development of UAV industry is booming,UAV system also shows great value and strong advantages in many applications.Among them,search and rescue mission is of great potential for that UAVs could overcome many difficulties traditional ways faced such as personnel safety and cost.It effectively enlarges the search area and speeds up the search process.In such system,the image processing part is its core concern,through multispectral and visible light image and LIDAR;with denoising,deblurring,image registration and image recognition,the efficiency of UAV search and rescue system could be greatly improved.However,UAVs’ aerial image processing and transmission system faces many chanllenges: 1.The high mobility of the UAVS: UAVs moves fast short time window of UAVs’ photo shooting which causes motion blur of the image;2.High resolution requirement: for the search and rescues mission demands the identification of specific object,the resolution needs to be high enough to support such mission;3.Rapid topology changes: UAV’s 3-D movement causes rapid topology changes and lead to difficulties in UAV’s image transmission.Our study reviewed the research status of aerial image processing,aerial image registration and ad-hoc network of UAVs.And with the applying of deep learning tools,improving of the ad-hoc network of UAVs(or Flying Ad-hoc NETwork,FANET),this dissertation did an in-depth study of the cooperated UAV search and rescue system and tried to solve the technical difficulties mentioned above.The main work of our study is listed below:1.The traditional deblurring method solve the problem through the estimation of the blur kernel,which cannot reflect the real-world situation of the UAVs.In this dissertation we proposed to apply Convolutional Neural Networks(CNN)as the deblurring tool.More over,a well-trained CNN network shows better performance compared with traditional method.2.As the CNN faces exploding gradients problem when a large number of layers are adopted,we used a layered training method and a res-net network structure to avoid the deterioration caused by batch normalization,etc.3.It is found that applying CNN to aerial image faces difficulties in getting clear reference images regarding the blurred training image set.In order to solve that,we proposed a new method to collect the traning set which is making UAVs taking pictures while they stay still and then start moving with continuous shooting to create the training set;the still images taking during such process are taken as the reference clear image and the continuous shooting blurred images taken as the training set to train the CNN.4.Traditional image registration method like Scale-Invariant Feature Transform(SIFT),shows satisfied result.However,it suffers from large computation cost,and hard to implement in mobile platform or used in real-time tasks.We proposed to use CNN to work with SIFT— using SIFT to find key points,and then using CNN to do the registration based on those points.Such method combined the high accuracy of the SIFT with high efficiency of trained CNN.At the same time,CNN could be trained offline through SIFT.The CNN is used to replaced two key steps of the SIFT—the determination of key points’ direction and generation of feature descriptor which consume most of the computation power,the result showed our method greatly reduced the computation time.5.Flying ad-hoc network is another important issue in UAV search and rescue system in order to transmit the aerial image.Ad-hoc network faces the time sensitive topology and the void problem when applying in UAVs.For that the UAV search and rescue system we proposed features its coordination of UAVs,we could use the known trajectory of UAVs in such system to improve the FANET protocol.We uses the trajectory to improve the Hello message sending process and the Multi Point Relay(MPR)selection process of the Optimized Link State Routing(OLSR)protocol to solve the outdated topology problem.6.To solve the void problem FANET faces under sparse distribution,we designed a new routing process utilizing the trajectory information and taking the store-forward(ferry like)process into account to solve the void problem under sparse distribution FANET.Major breakthroughs done by this dissertation were:1.Designed a method to get clear reference images regarding the blurred training image set for UAV aerial image deblurring.Based on the training set’s characters and CNN’s structure,selected and improved the loss function to train the model.2.Combining the SIFT algorithm and deep learning method,applied CNN in it and designed methods to train the model.3.Used the trajectory to improve the Hello packet related process and the MPR selection process of the OLSR protocol to solve the outdated topology problem,designed a new route-finding method for the protocol.Based on the above improvements,studies done in the dissertation show 0.92 d B improvement over the PSNR result,and clearer visual effect regarding the deblurring task comparing with the CVPR published methods;registration algorithm saved over 80% time in computation and the FANET simulation gives over 40 s reduction in end-to-end delay while 30% improvement regarding packet delivery ratio.The performance of the processing and transmission system of UAV was improved over all.