Heterogeneous sensor networks are increasingly deployed to support users in the field requiring many different kinds of sensing tasks. There may be multiple alternative kinds of sensors suitable for a given task. Sensing tasks might compete for the exclusive usage of available sensors. Such an environment is highly dynamic with users moving and generating tasks at different rates. Users typically lack the time and expertise to manually decide which are the best sensors for their tasks. We need therefore to design a distributed system to automatically allocate sensors to tasks. We formalize this problem as Multi-Sensor Task Allocation (MSTA) and show that the heterogeneity of sensors and tasks requires knowledge-based sensor-task matching. We extend a pre-existent well known coalition formation protocol to propose a novel layered distributed system which by using qualitative and quantitative measures provides allocation flexibility. We demonstrate that it is feasible to perform the knowledge-based sensor-task matching on a user's device by presenting a proof-of-concept mobile app which allows a user in the field to interact with the system. We run simulations to demonstrate that our architecture is scalable and that the allocation quality improves by allowing preemption of sensing resources from ongoing tasks and a reallocation mechanism.