Investigating passenger demand is essential for the taxicab business. Existing solutions are typically based on dated and inaccurate offline data collected by manual investigations. To address this issue, we propose Dmodel, using roving taxicabs as real-time mobile sensors to (i) infer passenger arriving moments by interactions of vacant taxicabs, and (ii) infer passenger demand by a customized online training with both historical and real-time data. Such huge taxicab data (almost 1TB per year) pose a big data challenge. To address this challenge, model employs a novel parameter called pickup pattern (accounts for various real-world logical information, e.g., bad weather) to increase the inference accuracy. We evaluate Dmodel with a real-world 450 GB dataset of 14, 000 taxicabs, and results show that compared to the ground truth, Dmodel achieves a 76% accuracy on the demand inference and outperforms a statistical model by 39%.