Zhendong Xu

Deep learning based methods have been increasingly applied to wireless localization and sensing. However, many existing methods suffer from poor generalization, which limits their adaptability to unseen environments. Additionally, retraining these models demands a substantial amount of labeled data, which is both time-consuming and labor-intensive. To address these challenges, we propose GenLoc, a generalization framework that enables accurate mitigation of ranging and angle estimation errors in unseen environments via feature disentanglement, without the need for additional data collection and online training. The key lies in learning fine-grained domain-invariant representations that are first extracted by minimizing feature distribution discrepancies across various domains and then refined through domain-invariant and domain-specific feature decoupling. Extensive experiments were conducted on real-world datasets, covering five distinct scenarios, each with five different obstacles, and a thorough comparison with several existing approaches illustrates that GenLoc improves average distance and angle estimation accuracy by more than 22% and 28%, respectively, while others exhibit negligible improvements or significant degradation.

Ranging error mitigation is key to improving indoor localization accuracy, but diverse environments challenge algorithm generalization. This letter proposes a multi-source domain adaptation framework for ranging error mitigation, where a weighted maximum mean discrepancy (MMD) is design to dynamically balances domain similarity and regression accuracy during feature alignment. By dynamically aligning domain mappings, the proposed method reduces negative transfer effects caused by varying discrepancies between different source and target domains. Experimental results show that the proposed network outperforms existing deep learning approaches in reducing ranging error across diverse indoor scenarios.

Deep learning methods have been widely applied in indoor localization to mitigate the impact of non-line-of-sight (NLOS) and multipath effects, achieving high-precision results. However, these methods have to integrate large amounts of data for training and often struggle to adapt to new and unseen scenarios. This letter proposes a domain generalization method based on domain adversarial training (DAT) and mutual information (MI) estimation, where cross-environment generalization localization is achieved by extracting domain-invariant features, eliminating the need for training data from new scenes. Experimental results demonstrate that the proposed method outperforms other deep learning approaches in terms of localization accuracy across various indoor environments.

The emerging B5G and 6G applications have brought forth the need for high-precision indoor localization. However, the complexity of indoor environments poses significant challenges to this goal, particularly due to the presence of non-line-of-sight (NLOS) conditions and multipath effects. This letter proposes an attention-based positioning network (ABPN) that exploits fine-grained features from MIMO channel state information (CSI) by spatial attention to combat the limited receptive field of traditional convolutional neural networks (CNNs) as well as channel attention to discriminate the importance of different wireless channels. Extensive experiments, conducted on two real-world datasets, demonstrate that the proposed ABPN outperforms the popular PirnatEco, AAresCNN, MIMOnet and CLnet with an average localization accuracy improvement of over 50%.

Crowdsourcing dramatically benefits WiFi fingerprinting localization in reducing the costs of collecting received signal strength (RSS) data during offline site survey and has gained much attention in the literature. This article proposes a deep-learning-based indoor positioning system (IPS), termed SeqIPS, to sufficiently exploit the available information in the crowdsourced sequential RSS data and inertial measurement unit (IMU) data. However, there exist the following three challenges: 1) the relatively large label noises of crowdsourced RSS data; 2) the unavailability of the labels of crowdsourced IMU data; and 3) the incorporation of the knowledge in IMU data into the localization model using RSS measurement as inputs during online localization. To this end, a co-teaching network is developed to effectively extract spatiotemporal features from sequential RSS data and meanwhile alleviate the influence of label noises. Also, a novel loss function involving IMU data is defined to impose spatial penalties, so as to further refine the localization model. Moreover, a domain adaptation module is included to effectively label the crowdsourced IMU data. Extensive experiments are conducted in a real scenario and show that SeqIPS can achieve an average localization error of 3.37 m, outperforming both traditional methods and recent deep-learning-based methods by 18.4% at least. In summary, the novelty of SeqIPS is that extra crowdsourced IMU data is exploited to refine the localization model during offline training, while only sequential RSS data is required as inputs during online localization, such that the system accuracy, simplicity, and costs are reasonably balanced.

In this letter, a novel localization scheme is reported to improve WiFi fingerprint-based localization via dimensionality reduction and an adaptive fingerprint matching algorithm. Specifically, an experimental analysis is firstly carried out to investigate the quality of radio maps. Then, the advanced uniform manifold approximation and projection (UMAP) algorithm is applied to refine the mapping relations from a fingerprint space to a location space in a supervised dimensionality reduction approach, which significantly improves the quality of radio maps. In addition, to further enhance the localization accuracy, boundary aware KNN (B-KNN) is proposed to adaptively deal with fingerprint matching according to inner area or boundary area. Extensive experiments are conducted in a real scenario with the area of nearly $1000~m^{2}$ , and a thorough comparison with several existing popular methods illustrates that the proposed scheme can significantly improve the localization accuracy on average by 33.3%.

Training an accurate and up-to-date radio map has always been a primary concern for implementing a WiFi fingerprint-based localization system. This paper presents a novel radio map learning scheme for online learning a set of kernel density functions, which function like a traditional radio map in WiFi fingerprint-based localization systems. To be specific, each kernel density function corresponds to one specific access point (AP) and predicts the probability of having a received signal strength (RSS) measurement from this AP at any position. In the offline phase, a set of coarse-grained kernel density functions are initially established by the Multivariate Kernel Density Estimation (MKDE) through a fast site survey. In the online phase, these kernel density functions are recursively refined through Incremental-MKDE (IMKDE) using freshly crowdsourced fingerprints, which essentially fuses the historical features in the previous kernel density functions and up-to-date features in the current fingerprints. A compression technique is applied to improve the computation and storage efficiency. In addition, pedestrian dead reckoning (PDR) is leveraged to calibrate the location labels of crowdsourced fingerprints. Extensive experiments are carried out in a real scenario of nearly 1000 $m^2$ for five months, and a comparison made between the IMKDE based method and several existing popular solutions confirms the superiority of the proposed IMKDE in terms of both computational complexity and localization accuracy.