Human Activities Recognition (HAR) is one form of video classification where the dataset contains several forms of Human Activities and the model predicts which class they belong to, using some traditional based or deep learning-based architectures. Here in this research, UCF-101 dataset is used which consists of total of 101 classes or human activities like—swimming, dancing, singing, etc. The Computer Vision-based Transformer (encoder-only model) with DenseNet121 as a feature extractor is used here as a deep learning model to train the data and classify them accordingly. An accuracy of 92.41% is achieved with this model on this UCF-101 dataset with proper optimization and hyperparameter tuning. This model gives better results than many state-of-the-art models and other traditional approaches used in this research. This model works better with more amount of organized diverse data and proper hyperparameter tuning.