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, WiFi-based HAR has gained significant attention due to its non-intrusive 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 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 function. This function dynamically learns to balance between traditional normalization and the Dynamic Tanh operation, thereby maintaining training stability and avoiding gradient vanishing issues often encountered when using Dynamic Tanh alone. Extensive experiments on two benchmark datasets demonstrate that the proposed TransHAR framework achieves state-of-the-art performance in both in-domain and cross-domain HAR tasks with only 0.17M 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 WiFi-based human activity sensing.

WiFi Indoor localization plays a crucial role in an emerging application domain for tracking indoor people, however, the serious issue is that the WiFi signals from access points (APs) vary greatly over time and the deployment structure of APs may be changed, for example, some APs are replaced or removed over time, which cause localization accuracy reduced. To solve this problem, this paper presents WiLoc, a Long-term WiFi localization with Lightweight Siamese Neural Network. This method introduces a Siamese neural encoder-based framework to learn the similarity between three inputs, where the Siamese network only consists of three linear layers without any convolutional layer or transformer. The triplet loss function is utilized to supervise the training of the feature encoder. Then, the encodings from this encoder are input to K-Nearest Neighbors (KNN) to predict the user’s positions. Extensive experiments on the UJI dataset, show the proposed WiLoc can effectively relieve the degradation of localization accuracy over time compared to the state-of-the-art algorithms, the degradation is reduced from 51% to 12.1%, and the average localization error is 2.06 m.

A Convolutional neural network (CNN) has emerged as a widely used approach to computer vision tasks, including object classification and detection tasks. The high requirement for the model to be more computationally efficient on lower information and communication technology (ICT) resource, e.g., mobile terminals can benefit from model distillation. However, most existing distillation methods suffer from a significant accuracy reduction, which requires a large number of pre-training models or doesn't make good use of the more of the network information, e.g., in the middle layers, during the distillation. In this paper, we study how knowledge about traffic signs recognition could be transferred to smaller models by distillation while cutting channels. We present an optimized object detection network, which uses a Region Proposal Network (RPN) weighted loss and hard-soft distribution-wise distillation loss for structural differences between teacher and student networks. We validate the network on multiple real-world datasets, the experiments demonstrate that the classification accuracy can be improved by 9 % with about 16 times parameter reduction while the detection network performance could be increased by 10.6% using an optimized object detection network.