18 July 2022 in IEEE Transactions on Human-Machine Systems
In this article, we present a work using a smartphone with an off-the-shelf WiFi router for human activity recognition with various scales. The router serves as a hotspot for transmitting WiFi packets. The smartphone is configured with customized firmware and developed software for capturing WiFi channel state information (CSI) data. We extract the features from the CSI data associated with specific human activities, and utilize the features to classify the activities using machine learning models. To evaluate the system performance, we test 20 types of human activities with different scales including seven small motions, four medium motions, and nine big motions. We recruit 60 participants and spend 140 hours for data collection at various experimental settings, and have 36 000 data points collected in total. Furthermore, for comparison, we adopt three distinct machine learning models, including convolutional neural networks (CNNs), decision tree, and long short-term memory. The results demonstrate that our system can predict these human activities with an overall accuracy of 97.25%. Specifically, our system achieves a mean accuracy of 97.57% for recognizing small-scale motions that are particularly useful for gesture recognition. We then consider the adaptability of the machine learning algorithms in classifying the motions, where CNN achieves the best predicting accuracy. As a result, our system enables human activity recognition in a more ubiquitous and mobile fashion that can potentially enhance a wide range of applications such as gesture control, sign language recognition, etc.
20 May 2025 in Discover Computing
The existing SLAM (Simultaneous Localization and Mapping) systems often suffer from increased positioning errors, inaccurate map construction, and challenges in real-time multimodal sensor data fusion in dynamic environments. This paper proposes enhancements to the SLAM system using nonlinear optimal filtering and deep learning to improve adaptability in such conditions. The study employs an Unscented Kalman Filter (UKF) for nonlinear state estimation, while deep feature extraction of environmental images is conducted via Convolutional Neural Networks (CNN). Semantic edge detection integrates Fully Convolutional Networks (FCN) and Canny edge detection techniques. The Extended Kalman Filter (EKF) is utilized for multimodal data fusion to optimize positioning accuracy across vision, lidar, and inertial measurement unit (IMU) sensors. Real-time motion estimation is achieved through an event-based camera combined with an optical flow algorithm, enhancing speed and accuracy in dynamic scenes. Experimental results demonstrate that the proposed SLAM system achieves an absolute trajectory error (ATE) as low as 0.067 m across datasets, with over 90% overlap in constructed maps. The system's average frame processing time is under 90 ms, and it responds within an average of 0.035 s during event-based camera experiments. These results outperform other mainstream SLAM systems, confirming that nonlinear optimal filtering and deep learning significantly enhance positioning accuracy, mapping quality, and real-time performance in complex environments. This study primarily employs simulation-based experiments to validate the proposed SLAM system's performance in dynamic hospital environments. While the results demonstrate high accuracy (e.g., an absolute trajectory error of 0.07 m), future work will include field experiments to further verify the system's robustness in real-world applications.
2 April 2025 in Discover Applied Sciences
Under complex lighting conditions, hospital logistics robots are facing serious challenges in positioning and navigation.The traditional ORB (Oriented FAST and rotated BRIEF) algorithm often has problems such as unstable feature point extraction, poor positioning accuracy, and long navigation path planning time in environments with large lighting changes, which greatly affects the robot's navigation efficiency and accuracy. In this paper, an improved ORB-SLAM3 algorithm is used to improve the navigation performance of the robot in a hospital environment with complex lighting. The improved ORB-SLAM3 algorithm extracts feature points of different scales by constructing an image pyramid to ensure that effective visual information can be stably obtained under different lighting conditions; the strategy of combining adaptive threshold and dual threshold methods is used to optimize the accuracy of feature point extraction, and the feature points are efficiently managed through a four-tree model to ensure the uniform distribution of feature points. These measures significantly improve the quality and matching accuracy of feature point extraction; the algorithm also combines visual and inertial information closely by combining inertial measurement unit data, further enhancing the positioning accuracy and system stability.The experimental results show that in the environment of rapid light changes, the average positioning error, navigation map construction integrity and navigation planning time of the improved ORB-SLAM3 algorithm are 0.26 m, 80% and 7.26 ms, respectively.This achievement not only effectively improves the navigation efficiency of the hospital logistics robot, but also enhances its stability and reliability in an environment with severe lighting changes. It is of great practical significance to improve the level of intelligence of the hospital's internal logistics system.
27 November 2025 in Complex & Intelligent Systems
The core challenge in remote sensing semantic segmentation lies in balancing local detail extraction with global contextual modeling. Existing methods often suffer from limited local receptive fields and inadequate global semantic interaction, leading to the attenuation of small-object features during hierarchical downsampling and fusion. To address these issues, we propose STRNet, a heterogeneous dual-stream encoder–decoder network that integrates a ResNet50 primary encoder and a Swin Transformer auxiliary encoder. This architecture jointly enhances local detail representation and global dependency modeling. Specifically, we design a feature refinement module (FRM) within the ResNet50 encoder to dynamically emphasize task-relevant features via channel attention, improving the recognition of small targets and complex regions. Meanwhile, the swin transformer branch incorporates a detail-preserving compression module (DPCM), which combines deformable convolutions and multi-scale pooling to optimize patch-token downsampling, thereby mitigating spatial detail loss during self-attention encoding. To effectively integrate multi-level features from both branches, we introduce a feature fusion module (FFM) that employs cross-branch channel-attention weighting and multi-scale alignment. The decoder adopts a progressive skip-connection strategy, fusing low-level spatial details with high-level semantics to construct a multi-resolution decoding pyramid for refined segmentation. Extensive experiments conducted on the ISPRS Vaihingen,ISPRS Potsdam, and LoveDA datasets demonstrate that STRNet significantly improves the segmentation accuracy of small objects and achieves competitive performance across all categories, validating its effectiveness for remote sensing semantic segmentation tasks. The code is available at https://github.com/Z09157/STRNet.
6 May 2025 in Discover Artificial Intelligence
Facing Gaussian noise from severe rain and haze, we employ multi-scale YOLO v10 for obstacle detection amidst high noise and DDPG (Deep Deterministic Policy Gradient) for enhanced path planning. YOLO v10 is used to annotate obstacles, and different input scales are set to deal with Gaussian noise. Recognition accuracy is optimized using cross entropy loss function. The DDPG algorithm is applied to update the path planning strategy, define the state space of robot position, velocity, and Laser Radar (LiDAR) readings, and output possible movement directions and velocities as action outputs. A reward function is designed to motivate the agent to learn. The weighted average method is used to fuse the visual information of YOLO v10 with LiDAR data. The robot’s status and surrounding environment are tracked through real-time feedback mechanism. The path planning is dynamically adjusted, and an intelligent navigation system is built. The research results show that the average recognition rate of the system for obstacles is 0.95, and the success rate of avoidance under extreme weather conditions of heavy rainfall is 88.0%. The system can effectively cope with various Gaussian noise environments and achieve precise path planning decisions.