Computationally Efficient Direct Position Determination Using Channel State Information

With the vigorous development of the Integrated Sensing and Communications(ISAC),high-precision positioning based on communication devices has become increasingly important for the accurate realization of advanced applications such as health monitoring,virtual reality and autonomous driving,etc.Among the various methods of positioning using communication devices,channel state information(CSI)stands out due to its ability to model location information explicitly,thereby providing refined positioning results.Therefore,research on positioning methods based on CSI is currently the hottest topic in the field of integrated sensing and communication.For CSI-based positioning problems,the direct position determination(DPD)technique has an advantage in the accuracy of position estimation,since traditional two-step positioning methods neglect the crucial constraint that intermediate parameters correspond to the same source.However,the shortcomings of conventional DPD methods are also quite apparent.The reason is that traditional DPD methods often estimate positions through grid search.On the one hand,the computational resources of communication systems are limited and cannot meet the requirement of dense grid points,leading to inevitable quantization errors.On the other hand,in the large-scale scenario,exhaustive search for maximum likelihood(ML)DPD of multiple target positions is inefficient and practically difficult to implement.To address the above issues,this dissertation explores computationally efficient DPD methods using CSI from two perspectives: iterative implementation and reducing gridpoint based calculation.The research can be summarized by chapters as follows: the fast DPD algorithm for the single source,grid-density-independent DPD algorithms in the presence of high-dimensional unknown variables,and the DPD algorithm based on dimension-reduced data.The main contributions include:1.To solve the problem that conventional single-source DPD search methods cannot process in real-time,this dissertation proposes a fast single source ML DPD algorithm(FSML-DPD)based on fast Fourier Transform(FFT).The position spectrum of FSMLDPD is obtained based on combining the FFT output of CSI,instead of directly generating it through exhaustive search.Then a large amount of redundant and repetitive computation is eliminated.Thus,compared to conventional single-source DPD methods,the computational complexity of the proposed algorithm is greatly reduced,while still maintaining consistent positioning accuracy.2.To efficiently perform DPD on a single stationary user equipment(UE)with moving receivers in scenarios where transmission frequency offset exists.This dissertation uses Expectation Maximization(EM)technique to iteratively estimate the target position by considering the source-receiver vector as the latent variable.Combining Laplace approximation and polar coordinate system,the algorithm transforms multi-dimensional search in conventional DPD into estimation of multiple one-dimensional polar angles,which prominently reduce the computational complexity,Moreover,the position update has the close-form solution.Although the estimation of transmission frequency offset is still required,the estimation process can be simultaneously used to detect whether the iteration is trapped by the local optimum and to escape from the local maximum by correcting the estimation method,contributing to the presence of nuisance parameters be advantageous.Therefore,although the proposed modified EM-based DPD algorithm(MEMDPD)is iteratively processing the non-convex cost function in nature,it can still obtain the global optima with any initial point.3.For the ML DPD problem in the multi-source scenario,conventional exhaustive search is computationally impractical.Based on the Pincus theorem,the DPD cost function can be transformed into an expectation computation function,so that position estimation is obtained by solving the mean of samples,and samples can be generated for each source separately through importance sampling(IS)of the original DPD cost function.Therefore,the computational complexity of the IS-based DPD algorithm(IS-DPD)is significantly reduced compared to exhaustive search,which can further be implemented in parallel.However,the existing IS-DPD methods uses an importance function that simply considers all target sources to be completely independent,without considering the strong coupling relationship when the sources are close.Therefore,this dissertation discusses a modified IS-based DPD algorithm(MIS-DPD)based on the improved adaptive importance function.The proposed algorithm classifies different sources according to their distances,and constructs an specific importance function for close-range sources considering the coupling relationship between them.Hence,proposed MIS-DPD is more robust and universal than conventional IS-DPD.4.In scenarios where communication resources are extremely limited,the CSI transmitted to the central processing station is often reduced in data dimension.First,in order to perform beamspace dimension reduction on CSI,only the optimal beam CSI containing the direct path is selected for subsequent processing.Therefore,this dissertation introduces an optimal beam selection method based on the first path has the minimum delay,which is more robust than the method based on beam power,especially in harsh indoor environments where the direct path is weak.After optimal beam selection,considering further reducing the data dimension by one-bit quantizing CSI.For the ML DPD problem of one-bit CSI,direct processing of ML cost function to estimate position requires extra iterative estimation of attenuation factors using gradient methods.However,here are complex integrals and exponential operations in gradient calculation.Thus,existing one-bit ML DPD methods have a large amount of redundant computation.This dissertation discusses a Majorization-Minimization(MM)-based one-bit DPD algorithm(MO-DPD),which iteratively solves the ML cost function.In each iteration,there is a closed-form solution for estimating redundant parameters,avoiding complex gradient calculations.This contributes to MO-DPD having much lower computational complexity compared to current one-bit ML DPD methods,while still maintaining accuracy.This dissertation exhibits the computationally efficient performance of the proposed algorithms through theoretical analysis and experiment verification,and also demonstrates the feasibility of CSI-based DPD algorithms in ensuring high positioning accuracy while maintaining low computational complexity.