An integrated YOLOv11 action recognition and vision-language model framework for intelligent kindergarten safety monitoring

A Multi-Modal Deep Learning Framework for Assistive Vision: Software-Based Design and Real-Time Validation

Martial Arts Routine Action Recognition and Scoring Using YOLOv8 and OpenPose

Martial arts routines are highly dynamic, multi-pose, and fast-paced, posing significant challenges to automated recognition and scoring. Such complex spatiotemporal characteristics are also representative of intelligent sensing tasks in advanced electromagnetic-aware environments, where robust human motion perception is essential for integrated visual and wireless monitoring systems. This paper proposes a joint framework combining YOLOv8 and time-optimized OpenPose to mitigate pose estimation jitter and detection inaccuracies caused by rapid motion and occlusion in human motion analysis. YOLOv8 first performs high-precision human target detection using bounding boxes to define regions of interest, after which the cropped images are processed by OpenPose to extract 17 keypoints, with IoU matching ensuring cross-frame identity association for temporal consistency. The keypoint sequence is modeled as a multi-channel time series and refined through a bidirectional LSTM network to predict smooth pose trajectories. The optimized keypoints are further used to calculate joint angles and movement velocities, which are integrated with dynamic thresholds for motion segmentation. DTW-based alignment and similarity matching with a standard motion library are subsequently employed to evaluate posture accuracy, motion amplitude, and rhythmic consistency, producing a comprehensive scoring result. Experimental results demonstrate average mAP@0.5 values of 0.863–0.942 (FPS 115.2–124.3), average PCK values of 88.7%–90.3% under occlusion (average MPJPE 48.2–51.6 pixels), and Pearson correlation coefficients of 0.925–0.941 for scoring consistency across diverse martial arts routines. The proposed framework provides an effective solution for intelligent motion analysis and offers technical reference for multimodal perception and dynamic scene understanding in advanced electromagnetic sensing applications.

Real-time multi-task learning with cross-task attention and edge supervision for autonomous driving perception

AI-Based Object Detection and Smart Navigation System for the Visually Impaired

EWFootstep 1.0: footstep data for early warning and surveillance systems

Due to the limitations of visual surveillance systems, such as line-of-sight, obtrusiveness, and high power requirements, audio-based surveillance has gained significant traction in security and forensic applications. Among these, footstep-based audio analysis has emerged as a promising and non-intrusive approach for monitoring and threat detection. This paper introduces EWFootstep 1.0, a novel dataset comprising recordings from 176 subjects collected under real-world environmental conditions, distinguishing footstep acoustic signatures of single and multiple individuals across forests, roads, and indoor settings. To validate the dataset, we perform time and frequency domain analyses, and implement machine learning (ML) and convolutional neural network (CNN) based baseline models. Feature separability is visualized using t-SNE and quantified using Davies-Bouldin Index. Our work bridges the gap in footstep-based security by providing a comprehensive dataset for machine learning applications in surveillance and crime scene investigation.

AN OPTIMIZED YOLOV10 FRAMEWORK FOR DUAL DETECTION OF SAFETY HELMETS AND LICENSE PLATES

This study suggests an enhanced safety helmet detection method based on YOLOv10 to solve the low detection accuracy of current algorithms for small objects and complicated settings in different situations.