Automatic detection of humans is an important part of many security and surveillance applications. Automatic processing diminishes the effect of human errors on the detection and allows extensive analysis after detection. Radar has certain advantages for human detection because of its all-weather and day and night capability, as well the fact it can function at a distance from the target. Radar-based systems to detect humans often consider the motion of these persons. Because of the micro-Doppler effect, human motion leaves a unique signature in a time-varying representation of the spectrum: the spectrogram, which can be used for the detection. Moreover, exploiting the micro-Doppler effect, allows to use low-cost and low-power consuming sensors, like a Continuous Wave (CW)-radar. The main objective of this research is to develop a algorithm that is capable of making a classification between human walking, human running and motion of other origin. Jointly, an estimation of motion parameters like velocity, height and phase of the gait cycle is performed. As approach, a particle filter implementation is chosen in combination with a model-based approach to analyze the human motion. The Thalmann model is used for human walking, while the Vignaud model is chosen for human running. A radar equipment model is developed to translate the kinematic positions and body dimensions first to a radar signal and thereafter to an estimated spectrogram. Based on the movements of some animals an estimated spectrum for the null-hypothesis was developed. Finally, the likelihood function for the particle filter was derived and the particle filter was implemented. The developed algorithm is tested both with simulated and with measured data. Because of its similarities to human walking, measurements of a walking duck were used to verify the null-hypothesis. For all inputs correct classification results were obtained. For human walking and running, the motion parameters were estimated correctly with satisfactory accuracy. An evaluation of the algorithm showed that Doppler resolution is very important to the performance of the algorithm. For Doppler resolutions poorer than 0.1 m/s the classification was performed wrongly and the Root Mean Squared Error (RMSE) of the estimated motion parameters increased. Time resolution was found to be non-decisive for the performance. Instead, the number of iterations that could be performed at a certain time resolution determined the performance of the algorithm.