In Mobile Crowdsensing (MCS), online task allocation is crucial as it effectively schedules and optimizes resource distribution, ensuring timely task completion and maximizing system performance in dynamic environments. However, conventional deep reinforcement learning methods have the limitations of low task allocation efficiency, unstable training, and insufficient experience sample utilization in MCS online task allocation. To address these problems, this paper proposes an online task allocation scheme called AD2QT (Attentive Dueling Double Q-Transformer). This scheme innovatively designs an attentive feature extraction network with a two-stage dynamic evaluation architecture based on Transformer to achieve adaptive modeling of environmental states. Additionally, AD2QT designs a deep reinforcement learning algorithm based on Double Q-Network and Priority Experience Replay to solve the problems of overestimating Q-values and low utilization of experience samples. Numerical experiments demonstrate that AD2QT outperforms other baseline algorithms in real-world scenarios, effectively adapts to various environments, and provides stable and superior solutions, further validating its effectiveness and reliability in online task allocation for MCS.