Recognizing sports activities from video frames using deformable convolution and adaptive multiscale features

Abstract

Automated techniques for evaluating sports activities inside dynamic frames are highly dependent on advanced sports analysis by smart machines. The monitoring of individuals and the discerning of athletic pursuits has several potential applications. Monitoring individuals, detecting unusual behavior, identifying medical issues, and tracking patients within healthcare facilities are examples of these applications. An assessment of the feasibility of integrating smart real-time monitoring systems across a variety of athletic environments is provided in this study. Motion and activity detection for recording sporting events has advanced due to the need for a large amount of both real-time and offline data. Through the use of deformable learning approaches, we extend conventional deep learning models to accurately detect and analyze human behavior in sports. Due to its robustness, efficiency, and statistical analysis, the system is a highly suitable option for advanced sports recording detection frameworks. It is essential for sports identification and administration to have a comprehensive understanding of action recognition. An accurate classification of human activities and athletic events can be achieved through the use of a hybrid deep learning framework presented in this study. Using innovative methodologies, we conduct cutting-edge research on action recognition that prioritizes users’ preferences and needs. It is possible to reduce the error rate to less than 3% by using the recommended structure and the three datasets mentioned above. It is 97.84% accurate for UCF-Sport, 97.75% accurate for UCF50, and 98.91% accurate for YouTube. The recommended optimized networks have been tested extensively compared to other models for recognizing athletic actions.