<p>This work investigates the benefits of employing a human-centric approach to video surveillance for robbery detection. We propose leveraging detected individuals in videos as input to a deep learning model, rather than directly extracting pixel information. Recent evidence has demonstrated the advantages of this approach in video action classification. Surveillance videos, particularly those captured by static cameras, are well-suited to this method. To this end, we introduce three self-supervised pretext tasks designed to incorporate human pose information during the training of a robbery detection model. Although our method did not surpass training from scratch in terms of AUC, it offers new insights into the potential and limitations of pose-based, self-supervised learning for video surveillance. Our findings contribute to a deeper understanding of how human-centric representations interact with current self-supervised learning techniques in the context of robbery detection.</p>