Research On Key Technologies Of Autonomous Navigation Based On Vision For UAV

With the advantages of light weight,small radar cross section,low operating cost,high flexibility and no crew safety problems,UAV has been widely used in meteorological detection,disaster monitoring,geological exploration and military reconnaissance,etc.As the core technology of UAV,high-precision autonomous navigation positioning is the key to ensure that UAV completes various tasks,successfully.However,in the flight environment of a complex UAV,GPS is vulnerable to environmental interference,and the cumulative error of inertial navigation will gradually increase with time,Which will lead to the navigation and positioning performance of the UAV is degraded or even invalid.Vision navigation as a new type of autonomous navigation can effectively make up for the shortcomings of GPS and inertial navigation.This article studies these deficiencies,and achieves the following innovative results:In recent years,a lot of researches on autonomous navigation of UAV based on vision have been carried out at home and abroad,and great achievements have been made,but there are still some problems to be solved as follows:1)It is not possible to quickly and accurately obtain high-resolution waypoint object images from Low Resolution(LR)images in the complex flight scene images;2)Threat targets and waypoints from the ground and in the air cannot be identified quickly and efficiently;3)The existing track planning algorithms cannot meet both real-time and optimal requirements.This article studies these deficiencies,and achieves the following innovative results:(1)Aiming at the unmanned aerial vehicle’s autonomous navigation process,the on-board image sensor is susceptible to the dynamic changes of the scene and the background factors,which makes it difficult to quickly and accurately obtain the waypoint target image from the aerial scene image.An algorithm for extracting the region of interest of aerial targets based on the attention mechanism of human eyes is proposed in this paper.Firstly,the complementarity between multiple features are used for establishing pre-attention and post-attention channels respectively to extract the saliency map of color,brightness,direction,position and global contrast characteristics of the scene image.Then,the feature saliency maps are normalized and fused to obtain the total saliency map that highlights the target area of interest.Finally,a local adaptive threshold segmentation algorithm is used to segment the total saliency map to obtain the waypoint target area of interest.Experimental results show that the algorithm proposed in this paper can quickly and accurately obtain the complete waypoint target region image from the scene image.(2)Aiming at the phenomenon that the target area image obtained by the UAV during autonomous navigation are blurred,high-frequency details are lost,and the spatial resolution is low,the super-resolution reconstruction algorithm for aerial imagery based on adaptive total variation regularization is proposed.Firstly,a hierarchical sub-pixel registration method combining time domain and frequency domain is used to perform sub-pixel registration on multiple frames of aerial target images;Then,the curvature difference is used for extracting the spatial structure information of the image to construct an adaptive regularization term instead of the traditional full variational regular term;Finally,the minimal optimization algorithm is used to iteratively optimize the adaptive full variational regularization model to obtain the final High-resolution image.Experimental results show that the proposed algorithm can improves the spatial resolution of the reconstructed image,which make the reconstructed target image contain more feature details.(3)Aiming at the fact that the airborne image sensor acquires the aerial ground waypoints and the threat target images are similar and indistinguishable during the autonomous navigation of UAV,which make it difficult to accurately identify it using the underlying feature recognition algorithm.The UAV target recognition algorithm based on transfer learning convolutional sparse autoencoder is proposed in this paper.Firstly the sparse autoencoder is used for performing unsupervised training on unlabeled data sets to obtain local features;Then using the convolutional neural network to perform knowledge transfer on aerial target image sets to extract global features.Finally,the global features of target images are sent to the classifier to complete the target object recognition.Experimental results show that the algorithm proposed in this paper can effectively improve the recognition accuracy.(4)In view of the fact that the existing dynamic trajectory planning algorithms cannot meet both the real-time nature and the status of optimal trajectory planning,the sparse A* track planning algorithm based on a cultural algorithm is proposed in this paper.Firstly,Based on the idea of digital map information fusion,the various threat sources encountered during the flight of UAV are equivalent to terrain obstacles according to their threat characteristics and strength,and superimposed on the digital map by its location to generate planning space with terrain information and various threat information.Then,making full use of the advantages of the sparse A* algorithm to quickly plan the track,and the dual-layer evolution optimization mechanism of the cultural algorithm for static and dynamic navigation Trace planning in the planned environment map.Experimental results show that the proposed method can effectively avoid sudden threats,ensure real-time track planning and take into account the optimal track.