Indoor Wi-Fi positioning based on Received Signal Strength Indicator (RSSI) has been widely used in all kinds of location-based services. However, the accuracy of positioning is susceptible to the sharp fluctuation of RSSI and the low sampling density. To address these difficulties, this paper proposes a novel fingerprint positioning method combining the virtual Access Point (AP) technique with Convolutional Neural Network (CNN) classification model. In our proposed method, a data augmentation method is designed to improve the positioning accuracy with few samples. The virtual AP location obtained by the distance ratio positioning method is utilized to adaptively modify the inputs of the CNN model. This new CNN model is used to determine the area of the samples, whereas a virtual AP method computes the accurate position of the samples. The experimental results show that the accuracy of area determination is up to 91%, and 95% of the positioning error is controlled within 2 meters. Compared with the existing positioning methods, the proposed method has a significant improvement in terms of positioning accuracy.

Due to the widespread application of location information, neural network localization method with the advantage of high localization accuracy has received significant interests in recent years. In this paper, we present two new neural network localization methods with time of arrival (TOA) measurements. In order to deal with three types of error about TOA measurements: measurement error, non-line of sight (NLOS) error and synchronization error, the proposed methods contain an offline training stage and an online localization stage. In offline stage, artificial neural network (ANN) or radial basis function (RBF) neural network is utilized to train the range measurements with the output of range errors rather than the position of mobile terminal (MT). Moreover, due to the unknown signal propagation condition whether it is line of sight (LOS) or NLOS propagation, k-mean clustering algorithm is used to classifying the range errors into different clusters. In online stage, the range errors are predicted and updated, then linear least square (LLS) algorithm with the adjusted range measurements is applied for the position estimate of MT. Comparing with ANN or RBF neural network localization methods, simulation results show that the proposed methods can effectively reduce the localization error, especially when the rtdtraining sample is not adequate. In addition, they are insensitive to measurement error, synchronization error and the distribution of NLOS error. Finally, the memory requirement and computational complexity about different algorithms are analyzed and compared.

While Wi-Fi Received Signal Strength Indicator (RSSI) fingerprinting has emerged as a prominent solution for indoor positioning, its accuracy remains hindered by labor-intensive data collection and environmental variability. To overcome these challenges, we propose a novel LCVAE-CNN methodology that integrates a Location-Conditioned Variational Autoencoder (LCVAE) and a multi-task Convolutional Neural Network (CNN) to enhance data quality and positioning performance. The LCVAE employs a dual-encoder architecture to augment RSSI fingerprints by jointly modeling signal features and spatial dependencies, introducing three key innovations: (1) dual-stream encoding that decouples RSSI and location processing for more effective feature learning, (2) a geospatial loss function that enforces topological consistency in the generated data, and (3) conditional data augmentation that preserves physical constraints of indoor spaces. The multi-task CNN then leverages shared feature extraction to jointly optimize classification and regression tasks, enabling efficient and accurate positioning. Extensive evaluations on the UJIIndoorLoc and Tampere datasets demonstrate the superiority of the LCVAE-CNN that achieves 98.80% floor classification accuracy with a Mean Positioning Error (MPE) of 6.79 meters on UJIIndoorLoc, whereas 97.22% accuracy with a MPE of 5.44 meters on the Tampere dataset. Compared to five state-of-the-art methods, it improves floor accuracy by at least 1.9% and reduces MPE by over 19%, while maintaining comparable computational overhead, thereby achieving superior accuracy-efficiency tradeoffs.

The range scale algorithm (RSA) proposed in 2004 is a constrained optimization problem with only three base stations (BSs). In this letter, by introducing a new variable to replace the square term, the quadratic objective function of the constrained optimization problem in RSA is turned into a linear objective function. Then the computational complexity is reduced and the performance is improved. Moreover, when the number of BSs is larger than three, we construct a new objective function and give the corresponding constraints. So the RSA is extended for more BSs and the performance of positioning is improved significantly. Simulation results demonstrate that our proposed algorithms improve the location accuracy over traditional algorithms in different Non-line-of-sight (NLOS) environments.