3D CNN Architecture for Video Scene Interpretation to Assist People With Visual Impairments

Ruiz Adarmes M.; Mauricio D.; Rivera L.; Huamani J.A.R. · IEEE Access (2026) · DOI: 10.1109/access.2026.3699268

Resumen

Video scene understanding in computer vision requires robust models capable of capturing both spatial and temporal information. Convolutional neural networks (CNNs) have traditionally excelled at extracting spatial features, while transformers have demonstrated superior performance in modeling sequential data and long-range dependencies. This paper proposes a 3D CNN architecture for video scene understanding in computer vision for the visually impaired, evaluating it based on performance metrics of accuracy and computational efficiency for action recognition in videos. The results reveal that the proposed model excels in modeling local spatial features and long-range dependencies, achieving an accuracy of 85.09% for validation. The article concludes with perspectives on future research directions, emphasizing the need for optimized architectures and improved datasets. Furthermore, experimental validation on an NVIDIA RTX A1000 GPU with 8 GB memory demonstrates that our model achieves a favorable accuracy-resource trade-off, requiring only 1240 MB of GPU memory and 92 ms inference time per video, outperforming Transformer-based methods in resource-constrained scenarios.

Instituciones: Universidad Nacional Mayor de San Marcos; Universidad Nacional de Ingeniería