Arterial blood pressure (ABP) is one of the most important vital signs for the assessment of the human cardiovascular system. ABP is mainly characterized by the systolic (SBP) and the diastolic blood pressure (DBP). Early detection of elevated ABP is of utmost importance to prevent manifestation of severe diseases such as stroke. In order to achieve regular self-monitoring of ABP, measurement devices should be as convenient as possible. This work investigates a novel approach based on RGB cameras to estimate ABP remotely. We applied imaging photoplethysmography (iPPG) for the extraction of the blood volume pulse (BVP) from 56 healthy subjects with induced elevated SBP. The BVP was then used as the input for a customized neural network (Vnet) operating in time and frequency domain. SBP and DBP were estimated based on 20-second segments. To evaluate our approach, nested cross validation was applied. We then calculated the Pearson correlation coefficient (r) to quantify model agreement for SBP and DBP estimation. For generalized ABP estimation, Vnet achieved r of 0.30 and 0.19 for SBP and DBP, respectively. In contrast, for personalized ABP estimation, r reached 0.71 and 0.51 for SBP and DBP. Our work demonstrates the potential of ABP estimation based on iPPG and delivers important insights into the relevance of optimal model selection from cross validation.