Research On Path Planning Method Of Interest Target Based On Deep Learning And Remote Sensing Image

Path planning of interest targets has great application value in the fields of traffic navigation,command and dispatch,and emergency rescue,etc.Conventional methods are generally implemented based on topological road data and interest target database.They rely on the current nature and integrity of existing road networks and target data.In some emergency situations such as disaster relief and battlefield maneuvering,existing road network data may not meet the requirements.Moreover,the data update period is relatively long,which will restrict the implementation of the action.Compared with road network data,the present situation of remote sensing data is strong and its acquisition method is rich.Therefore,remote sensing image can be used to automatically generate the road network data based on intelligent means,thereby realizing the emergency path planning without the support of road network database.This paper studies the interest target detection,road network segmentation and automatic topology construction involved in intelligent path planning of interest target in remote sensing image.The main contents and innovations are as follows:1.Aiming at the problem that interest target in remote sensing image is difficult to detect due to its small size and different scale,an improved SSD target detection algorithm is proposed.The algorithm uses a densely connected network as the feature extraction network in the SSD,and constructs a gold tower feature map between the densely connected modules for target detection.To verify the performance of the algorithm,an online sample data acquisition system is designed.2,574 images of aircraft and sports fields,mainly small and medium-sized targets,were collected as data sets.Experimental results show that the algorithm can still achieve good results without the support of migration learning.Compared with Faster R-CNN and R-FCN algorithms,the proposed algorithm increases mAP by 3.64% and 2.97%,respectively.In the small target detection,the mAP is increased by 7.4% and 5.15%,respectively.The training process is not easy to diverge,and it takes less time to detect target.2.Aiming at the problem that the existing deep learning semantic segmentation methods for road network segmentation are difficult to fully consider the relationship between pixels,a road network segmentation method combining semantic segmentation model and convolution condition random fields is proposed.The proposed method improves the existing Deeplabv3+ model.Firstly,the ReLU activation function in Deeplabv3+ is replaced with the SELU activation function to build SELU-Deeplabv3+.Secondly,based on the feature pyramid method,the low-level features are combined with high-level semantic information.Then the improved Deeplabv3+ is used as the semantic segmentation network,and the convolution condition random fields post-processing module is introduced at the back end of the network to realize structured prediction.This paper uses Massachusetts road dataset and DeepGlobe 2018 road dataset as road segmentation datasets.Experimental results show this method can achieve good segmentation accuracy.Recall at breakeven(Bt)in the Massachusetts road dataset reaches 90.89%,an increase of 0.83% and 0.42% compared to the Mnih's method and Saito's method,respectively.On the DeepGlobe 2018 road dataset,MIoU is increased by 4.65% and 2.45% compared with LinkNet34 and Topology Loss respectively.3.Overall implementation framework and process of fast path planning for interest target of remote sensing image with improved SSD and road segmentation algorithm are designed,and the key technologies are studied and analyzed.A wide range of target detection and road network segmentation are realized based on the meshing idea.Based on the idea of tile map,a large-scale road segmentation map can be indexed and stored,so as to realize fast access and seamless splicing of large-scale road segmentation data.According to the different requirements of target path planning task,the path planning method based on skeletonized road raster map and A* algorithm and the path planning method based on road topology map and Dijkstra algorithm are designed.In the process of constructing the road topology diagram,the expression of graph is designed,and a method of convolution combined with array index is proposed to realize the road node search and optimization.Experimental results show that the proposed path planning method can only use tens of seconds in a small area to realize path generation under the condition of remote sensing image.4.The interest target search and its path planning verification platform under the B/S architecture is designed and developed.Through interest target detection,vector road extraction and path planning experimental analysis,the stability and reliability of the platform are verified,which provides a more effective solution for the practical application of the algorithms.