Accurate and real-time outdoor localization and pose estimation are critical for various applications, including navigation, robotics, and augmented reality. Apple’s ARKit, a leading AR platform, employs visual–inertial odometry (VIO) and simultaneous localization and mapping (SLAM) algorithms to enable localization and pose estimation. However, ARKit-based systems face positional bias when the device’s camera is obscured, a frequent issue in dynamic or crowded environments. This paper presents a novel approach to mitigate this limitation by integrating position bias correction, context-aware localization, and human–computer interaction techniques into a cohesive interactive module group. The proposed system includes a navigation module, a positioning module, and a front-end rendering module that collaboratively optimize ARKit’s localization accuracy. Comprehensive evaluation across a variety of outdoor environments demonstrates the approach’s effectiveness in improving localization precision. This work contributes to enhancing ARKit-based systems, particularly in scenarios with limited visual input, thereby improving user experience and expanding the potential for outdoor localization applications. Experimental evaluations show that our method improves localization accuracy by up to 92.9% and reduces average positional error by more than 85% compared with baseline ARKit in occluded or crowded outdoor environments.
6 November 2025 in IEEE Transactions on Artificial Intelligence
Human Activity Recognition (HAR) plays a crucial role in intelligent healthcare, smart environments, and elderly monitoring. Traditional deep learning-based HAR methods often function as black-box models, limiting their interpretability. The recently proposed Kolmogorov-Arnold Network (KAN) utilizes explicit, mathematically defined basis functions, which clarify its operation and enhance interpretability. However, these methods still face challenges, such as slow training speed, high computational costs and suboptimal performance. Here we propose the Daily Activity Recognition with Optimized Wavelet-based KAN (DARKAN), a lightweight architecture that leverages wavelet decomposition to boost performance, and simplifies KAN structure to lower model parameters and computational complexity. Specifically, low- and high-frequency inertial measurement unit (IMU) signals are extracted by a wavelet transform, while time-domain features are incorporated to enrich feature representation. Subsequently, the B-Spline is replaced by the wavelet function as the activation function in KAN (wav-KAN), and the network depth of wav-KAN is reduced to two layers. Finally, the optimized wav-KAN is utilized to classify daily activities by fusing the extracted time-frequency features. Extensive experiments on three open-source datasets demonstrate that DARKAN outperforms stat-eof-the-art methods, achieving 98.82%, 97.11% and 98.57% in classification accuracy respectively while reducing the number of model parameters by 1.45× and FLOPs by 4×. The code is available at https://github.com/AnonysHar/HarKAN.
Human activity recognition (HAR) has emerged as a critical component in intent-aware, AI-driven Internet of Things (IoT) Communication systems, enabling context-aware responses in smart environments. Recently, Wi-Fi-based HAR has gained significant attention due to its nonintrusive nature, low deployment cost, and ability to preserve privacy. However, they face a major challenge across domains. To address this limitation, we propose a novel cross-domain HAR framework (called TransHAR) by introducing a lightweight and efficient transformer model. On the one hand, we design a feature representation block that processes the Wi-Fi channel frequency response (CFR) phase data to estimate Doppler shifts, capturing motion-related dynamics while remaining invariant to static, environment-specific structures, enhancing generalization across domains. On the other hand, we propose a lightweight transformer architecture, termed ResDyTFormer, which minimizes reliance on normalization layers by incorporating a novel residual dynamic Tanh (DyT) function. This function dynamically learns to balance between traditional normalization and the DyT operation, thereby maintaining training stability and avoiding gradient vanishing issues often encountered when using DyT alone. Extensive experiments on two benchmark datasets demonstrate that the proposed TransHAR framework achieves the state-of-the-art performance in both in-domain and cross-domain HAR tasks with only 0.17 M parameters. On the SHARP dataset, it attains an impressive 99.04% F1-score and 98.94% accuracy. On the 3DO dataset, it achieves 86.10% accuracy and 84.95% F1-score. These results highlight the potential of TransHAR as an efficient and scalable framework for real-world Wi-Fi-based human activity sensing.