Soft evidence sources play a critical role in social networks and similar settings, where subjective evidence and opinions are the norm. Study of opinion dynamics (including consensus and cluster formation) in these scenarios requires agent models that can capture the types of uncertainties and nuances characteristic of soft evidence (human-generated input, subjective evidence, etc.). To address the corresponding challenges, we employ a Dempster-Shafer (DS) belief theoretic agent model to explore opinion dynamics under bounded confidence. The proposed model further captures the notions of global affinity and the nature of persuasion of agents in social judgement theory. The paper develops several new results and these results regarding formation of clusters and consensus of agent opinions are verified with the aid of several numerical studies accompanied by bifurcation diagrams.