| In recent years,UAV has developed rapidly and is widely used in all walks of life.Compared with fixed wing UAV,rotor UAV has better portability,which makes it more widely used.Some professional rotary wing UAVs often carry more important professional equipment.If the rotary wing UAV has an emergency such as GPS failure during flight,it may damage the UAV and airborne equipment,and even cause economic losses and personal injury to the ground equipment and pedestrians.At present,the research on the flight safety emergency landing of rotor UAV is mainly considered in the control algorithm,in order to maintain the attitude and flight stability of UAV;The research on visual guided landing of rotor UAV is mainly to identify the safe landing sign with label information at the predetermined location to complete the landing.This thesis studies that the rotor UAV depends on machine vision to find the safe landing area and safe landing place in emergency,and solves the problem of visual landing location of rotor UAV in emergency.This thesis proposes to use the deep learning semantic segmentation network model to segment the aerial image collected by the airborne camera of the rotor UAV,and extract the safe landing area of the scene under the UAV.The UNET semantic segmentation model and ICG drone dataset are used to build the model and experiment,and then pruning optimization is improved based on the UNET model.The experimental results show that the improved network model increases the speed of network operation and reduces the size of network model parameters by about 60 MB.Under the same conditions,there is no decline in accuracy and other indicators between the optimized network model and the original network model,which ensures the operation speed and effect of the network.This thesis proposes a safe landing location algorithm based on deep learning semantic segmentation graph,including inscribed circle center method,improved inscribed circle center method and kernel density estimation method.Firstly,the object scene category suitable for rotor UAV and the safety of landing point are discussed,and then the model is used to select the safe landing place in the safe landing area of semantic segmentation map.The experimental results show that when the maximum inscribed circle center of the safe landing area is selected as the landing point by the inscribed circle center method,the effect is affected by the semantic segmentation model;The improved inner circle center method makes up for the shortcomings of the inner circle center method and has faster operation speed than the kernel density estimation method;The operation speed of kernel density estimation method is very sensitive to the area and shape of the safe landing area.When there are a large number of safe landing areas in the image,the operation speed of the model is too long.Finally,the collected aerial images of UAV are used to experiment,analyze and evaluate the whole location model,and the relevant original experimental data are given.The location accuracy of different models is more than 85%.In addition,the thesis also discusses the parameters such as the area size of the safe landing area required by different UAVs and the mobile landing strategy of UAVs.Finally,some suggestions on the application of machine vision in the emergency landing of rotor UAV are given.The relevant source code is disclosed in the appendix. |
