In recent years, falls among the elderly have become a critical healthcare concern worldwide. There is more and more research every day to find ways to detect falls early, hence reducing the associated risks and improving the quality of life. Many published datasets are used to create fall detection models. However, there is still room for improvement when the datasets often face an imbalance and a lack of diversity. To solve this problem, a new dataset has been built for the fall detection problem, namely ViFam in the hope of contributing to creating a better model for fall detection. In this work, the built dataset is used for two problems: fall detection and activity recognition. They are experimented to evaluate the performance of 6 models: CNN, ResNet, ViT, 3D-CNN, and Mixture of Experts with 5 and 6 experts. Ultimately, 3D-CNN achieved the best results, with all metrics reaching higher than 0.99 for both problems. We aim to publish this dataset to contribute to the recent research on the fall detection problem.