The relevance of this research is driven by the need to improve the accuracy of human action recognition systems in video by integrating audiovisual information, which compensates for the limitations of each individual modality. The aim of this work is to develop and investigate methods for late fusion of predictions from independently trained video and audio models for recognizing six classes of human actions (Archery, BalanceBeam, BasketballDunk, BoxingSpeedBag, FloorGymnastics, SoccerPenalty). The study employs 3D convolutional neural networks with Squeeze-and-Excitation blocks for video processing and a hybrid CNN-GRU architecture for audio analysis, along with seven prediction fusion methods: arithmetic mean, weighted average, majority voting, confidence-based fusion, stacking, product of probabilities, and the Borda count. Experiments demonstrate that the arithmetic mean and product of probabilities methods achieve the highest accuracy of 91.67%, outperforming the individual modalities by 2.33% for color and by 5.6% for grayscale video models. Confidence-based fusion exhibited the best results for grayscale models by automatically adapting the contribution of each modality based on the entropy of their predictions. Special attention is paid to the strict separation of training, validation, and test sets based on filming environments to prevent data leakage and to ensure objective evaluation of model generalization capability. The obtained results confirm the practical value of the proposed approach for video surveillance systems and multimedia platforms, offering an optimal balance between accuracy, computational complexity, and robustness to the absence of one modality. This research contributes to the development of multimodal machine learning methods and provides practical guidelines for selecting a fusion strategy depending on available computational resources and input data types.
