The main challenge in cuffless blood pressure (BP) monitoring lies in its inaccuracy during daily activities. Many previous works either lacked appropriate validation with sufficient intra-individual BP variations, or showed unsatisfactory estimation accuracy under activities. This study introduces a novel deep learning model called UTransBPNet, which aims to improve the estimation accuracy and tracking capability of intra-individual BP changes under activities. UTransBPNet combines the Unet architecture with self- and cross-attention from transformer to estimate BP waveforms, which are then further regressed to systolic (SBP) and diastolic BP (DBP) using additional fully connected layers. The model’s performance was thoroughly evaluated in three distinct datasets: one public dataset (Dataset_MIMIC) and two datasets during activities (Dataset_Drink and Dataset_Exercise). Under subject-independent validation, the model achieved mean absolute differences (MADs) of 4.38 mmHg for SBP and 2.25 mmHg for DBP in Dataset_MIMIC. Additionally, the tracking capability of intra-individual BP variations during activities were also significantly improved. In Dataset_Drink, Pearson’s correlation coefficients for SBP and DBP were 0.61±0.17 and 0.62±0.13, respectively. In Dataset_Exercise, the corresponding correlation coefficients were 0.82±0.11 for SBP and 0.72±0.18 for DBP. Furthermore, this study for the first time tested the generalization capability across different activities. The results showed that our model, with small-sized scenario-specific data for finetuning, exhibited a good cross-scenario generalization capability. However, it degraded significantly when there were differences in the BP distributions and variation patterns between the datasets. In conclusion, UTransBPNet shows promise as a deep learning model for accurate cuffless BP estimation and tracking BP changes during activities.