A Hybrid Deep Learning Model that Uses Spatiotemporal and Distinctive Features for Sports Behavior Recognition

Recognizing human actions in sports videos is a critical task in computer vision with applications in performance analysis, coaching support, and automated game understanding. Traditional machine learning and shallow vision models often struggle to capture the complex dynamics of real-world sports behaviors, especially when dealing with occlusions, fast motion, and varying camera perspectives. These limitations result in reduced accuracy and generalization capabilities, particularly in temporal action sequences. To overcome these challenges, a hybrid deep learning framework is proposed that combines convolutional neural networks (CNNs) with a TimeSformer module. The CNN component efficiently extracts spatial features from individual frames, capturing object details and player positions, while the TimeSformer models temporal relationships across sequences using self-attention, enabling the detection of motion patterns and action transitions. Preprocessing steps, including Gaussian filtering and min–max normalization, ensure clean and consistent frame inputs, while fixed-length video segmentation prepares data for temporal analysis. Additionally, the navel gazing optimization (NGO) algorithm is employed to fine-tune hyperparameters, such as learning rate, dropout rate, and model depth, optimizing the performance without manual intervention. The model is evaluated using a labeled sports video dataset covering actions like sprinting, passing, and jumping. Results demonstrate high accuracy, with a confusion matrix and Receiver Operating Characteristic (ROC) analysis confirming strong classification performance. The model achieves over 97% validation accuracy and shows minimal false positives and false negatives, indicating its robustness and reliability. The proposed approach significantly improves the recognition of sports behaviors by learning both spatial and temporal representations in a unified framework. The integration of NGO enhances model efficiency, making the method suitable for real-time or resource-constrained environments. This hybrid architecture offers a scalable and intelligent solution for action recognition in diverse sports applications.