The widespread application of human activity recognition (HAR), which serves as a bridge to narrow the gaps between healthcare and smart home, is now moving towards supervising individual health in a manner that will outsmart the human mind. This paper delves into the depths of HAR and was created to train machine learning classifiers for HAR based on the increasing number of professional annotations of activities in free-living settings. Deep learning, which is based on artificial neural networks, has demonstrated its effectiveness in many areas, such as virtual assistants, autonomous driving, image recognition and classification, and speech processing due to its ability to learn from vast quantities of data. In this study, we present HAR_Net, a two-stage pipeline for solving the problem of HAR from wearable sensor data. In the first stage, the 3D time-series sensor data is encoded into a 2D image using the concept of Gramian Angular Field (GAF). In contrast, in the second stage, the GAF images are classified using a customized Convolutional Neural Network (CNN), which achieves classification accuracies of 90.56, 92.60, and 90.4% on the HARTH, HARSense, and Mobile Health Human Behaviour Analysis datasets, respectively, outperforming many state-of-the-art HAR models.