Deep Learning Method For Single And Multiple Target Localization Based On Angle-of-Arrival Measurements

Passive positioning technology based on the angle of arrival has the advantages of high concealment,long detection distance,and simple operation.It has received more and more attention from scholars in the fields of military reconnaissance,indoor positioning,and wireless sensor networks.Different from traditional statistical parameter estimation methods,this thesis attempts to use a modern deep learning framework to solve the problem of single and multiple signal source localization based on the angle of arrival.It is expected to improve the accuracy of localization through empirical learning.In single-target localization tasks,most algorithms currently solve the target point coordinates based on statistical optimization.With the increment of the noise level,the localization performance of the algorithm drops dramatically.These algorithms generally lack strong generalization ability and unable to learn errors in the localization environment effectively.In the multi-target localization task,with the increase of the number of anchor points and target points,a large number of false intersections will appear in the direction line of the localization space.Effectively eliminating false target points is the key to improving the performance of multi-target localization.Existing multi-target localization algorithms generally have following shortcomings:?They cannot effectively remove false target points,and can only remove false target points with a few anchor points and target points.When the number of anchor points increases,false alarm rate will increase greatly.?They cannot effectively learn the error distribution of localization environment,indicating poor generalization ability.Aiming at the problems of the above single and multiple target localization methods,this thesis transforms the domain of the localization task to the image rendered by the intuitive geometric data and uses modern methods of computer vision and deep learning to get solutions.Specifically,this thesis treats single-target localization a visual target position regression tasks,and converts the multi-target localization tasks into semantic segmentation of the corresponding rendered image.so as to realize the use of deep neural network models to locate single and multiple signal sources with high accuracy.The main contents and innovations of this thesis are as follows:(1)In the single-target localization task,this thesis we proposed a new idea method to solve the target localization problem,a single-target fixed task is converted into a computer vision regression task by transforming the solution domain,and a sample generation algorithm is used to generate a data set in this thesis.The transformation of the solution domain and the sample generation algorithm are the main key parts of the single-target localization task.A localization convolutional neural network(LocNet)model is proposed to model examples.The complex representation of the LocNet model is used to learn the spatial distribution around target point in the environment to improve localization performance of the target.Simulation experiments show that at higher noise levels,the localization performance of our model exceeds the Cramer-Rao Lower Bound(CRLB)of the traditional least squares model,and has strong generalization ability.(2)In the multi-target localization task,this thesis transforms the multi-target fixed task into a computer vision semantic segmentation task by transforming the solution domain.According to our research,this thesis is the first application of a semantic segmentation neural network model to the field of multi-target positioning based on direction finding data.among them,the transformation of solution domain and example generation algorithm are the core content of multi-target localization.This thesis proposes a multi-target localization semantic segmentation network(MLocNet)model based on the encoding-decoding architecture to denoise and reconstruct the sample image space,thereby eliminating false target points and improving multi-target localization performance.Simulation experiments show that with the increase of the number of anchor points and target points,compared with other current multi-target localization algorithms,the proposed model can effectively eliminate false target points,with high coverage and low false alarm rate.Even when the number of anchor points is 20 and the number of target points is 2-20,the localization performance F₁ index of our model can sequentially exceed the localization performance of the baseline minimum distance method by 37%.