Abstract This paper presents an exercise movement recognition system based on 3D Convolutional Neural Networks (3D CNNs) to enhance intelligent exercise teaching. The author, Binbin Gong, addresses limitations of traditional instruction such as lack of real-time, personalized feedback by leveraging 3D CNNs to extract spatiotemporal features from exercise videos for accurate motion classification. The model architecture includes four convolutional layers, two... This paper presents an exercise movement recognition system based on 3D Convolutional Neural Networks (3D CNNs) to enhance intelligent exercise teaching. The author, Binbin Gong, addresses limitations of traditional instruction such as lack of real-time, personalized feedback by leveraging 3D CNNs to extract spatiotemporal features from exercise videos for accurate motion classification. The model architecture includes four convolutional layers, two max pooling layers, two fully connected layers, and a Softmax classifier. Evaluated on the MSRAction3D dataset, the system achieved an average recognition accuracy of 91%, with individual action accuracies ranging from 81% to 98%. The study reviews current intelligent technologies in exercise education, including pose estimation, VR, and deep learning methods like RNNs and CNNs, both internationally and domestically. Results demonstrate the model's strong potential to support instructors with real-time movement analysis and personalized student feedback, thereby improving teaching effectiveness and learner experience in physical education. Read More Read Less

Abstract The paper presents a 3D ball tracking framework tailored for complex team sports environments, such as basketball and football, where challenges like occlusion, small object size, and dynamic backgrounds hinder accurate motion analysis. The proposed method integrates a small object detection network based on multiscale features with the Faster R-CNN architecture to improve 2D ball detection across multiple camera views.... The paper presents a 3D ball tracking framework tailored for complex team sports environments, such as basketball and football, where challenges like occlusion, small object size, and dynamic backgrounds hinder accurate motion analysis. The proposed method integrates a small object detection network based on multiscale features with the Faster R-CNN architecture to improve 2D ball detection across multiple camera views. To address tracking drift and missed detections, the study enhances the ECO (Efficient Convolution Operators) tracking algorithm by incorporating cross view information via epipolar constraints and updating the tracking model using detection outputs. These refined 2D coordinates from synchronized cameras are then fused into 3D space using a triangulation algorithm, which computes accurate 3D ball positions while filtering out results with high reprojection errors. A simplified nonlinear motion model combined with a Kalman filter further smooths the 3D trajectory, yielding precise and stable tracking paths. Experiments on the APIDIS basketball dataset and self collected sports videos demonstrate significant improvements: 2D tracking accuracy rises from 0.67 to 0.77, and 3D tracking accuracy reaches 0.92 with the enhanced Kalman filter. The framework effectively addresses real-world issues such as occlusion, lighting variation, and camera calibration errors, providing a robust solution for sports video analysis. This approach supports advanced applications such as referee assistance, tactical evaluation, and performance analytics, highlighting its potential to advance intelligent systems in competitive sports.. Read More Read Less

Abstract The paper presents a method for correcting weld seam errors in intelligent robotic welding systems using three-dimensional (3D) visual localization technology. Authored by Jinbo Li from Henan Vocational College of Logistics, the study addresses limitations in conventional robotic welding, particularly the inability to detect and correct seam deviations in real-time. The proposed approach integrates 3D visual sensing (via laser scanning... The paper presents a method for correcting weld seam errors in intelligent robotic welding systems using three-dimensional (3D) visual localization technology. Authored by Jinbo Li from Henan Vocational College of Logistics, the study addresses limitations in conventional robotic welding, particularly the inability to detect and correct seam deviations in real-time. The proposed approach integrates 3D visual sensing (via laser scanning or structured light) with machine learning algorithms, such as Support Vector Machines and deep neural networks, to accurately identify and classify weld seam geometry from point cloud data. Key features like curvature, normal vectors, elevation statistics, and spatial density are extracted to enhance localization precision. The system enables real-time error correction by comparing actual seam positions with predefined paths, adjusting the robot’s trajectory accordingly. Experimental results demonstrate high accuracy, with a relative error of only 0.38% in distance measurement and endpoint positioning errors ranging from 2.5 to 4 mm, meeting industrial standards for crane beam welding. The paper validates the effectiveness of the algorithm through multiple tests, including Canny edge detection and dual-pixel scanning for depth analysis. The authors conclude that combining 3D visual localization with intelligent algorithms significantly improves welding quality, reduces defects, and enhances the adaptability and efficiency of robotic welding systems. Future work aims to refine classification models and further integrate deep learning for more robust seam tracking and correction in complex industrial environments. Read More Read Less