Human Activity Recognition (HAR) is a significant area of interest with diverse potential applications; however, the existing literature lacks a high-performance reference solution. This study aims to fill this gap and focuses on the domain of device-free HAR, investigating a little-explored vision-based approach for action representation: the Encoded Human Pose Image (EHPI). The research explores various architectural choices for EHPI generation and provides a performance baseline using the public MCAD dataset. Additionally, to overcome the difficulty of limited datasets in HAR in literature, as an additional contribution we introduce and share the UCBM-ELT dataset. It is an in-house realized dataset that focuses on atomic-level action analysis in more complex and realistic scenarios and aims to enhance model generalizability and robustness by incorporating intra-class variations.