In high-end, data-intensive embedded sensor applications (radar, optronics), the evolution of algorithms is limited by the computation platform capabilities. These platforms impose Size, Weight and Power (SWaP) restrictions on top of reliability, cost, security and (potentially hard) real-time constraints. Thus mostly static mapping methods are used, negating the system's adaptation capabilities. Through the study of several industrial use-cases, our work aims at mitigating the aforementioned limitations by introducing a low-latency dynamic resource management system derived from techniques used in large-scale systems such as cloud and grid environments. We expect that this approach will be able to guarantee non-functional properties of applications and provide Quality of Service (QoS) negotiation on heterogeneous platforms.