Use a filter that fits a low-order polynomial to your signal. I would recommend a recursive IIR filter (so that the computational complexity is independent of the impulse response duration). Set the impulse response duration to give you the best possible white noise attenuation, while maintaining a reasonable transient response. Tune the pass-band delay to give you the best possible white noise attenuation, while maintaining a reasonable filter latency. The filter outputs the (least-squares) fitted polynomial coefficients. So you can compute time derivatives from these coefficients. Most people will probably recommend a Kalman filter. If your models are static and if your sampling interval is constant then it will converge rapidly to the steady-state filter that I describe (which does not require covariance matrix manipulations, gain computation etc.).   

Thank you very much.