Human activity recognition (HAR) is critical in several domains, such as surveillance, human-computer interaction, and healthcare. By preprocessing the KTH Dataset with convolutional neural networks (CNN) and generative adversarial networks (GAN), particularly DCGAN (Deep Convolutional GAN), we offer a novel method in this study to enhance HAR. We preprocess the dataset by using Min-Max normalization and converting video frames to grayscale to achieve homogeneity. To refine the data even further, DCGAN is also employed. Next, we train our CNN model to recognize human actions using the preprocessed data. Experimental results show that compared to current methods, the recommended model greatly improves action recognition accuracy. Additionally, this model has exceptional performance in various activity areas. This study not only advances HAR approaches but also demonstrates how CNN and GAN methodology may be used to preprocess complicated datasets such as KTH.