We propose a new semi-automatic strategy for the localization of brachyther-apy seeds with transrectal ultrasound imaging. We formulate the problem as a rigid surface-to-image registration, where a geometric model is embedded in an external force field pointing towards the last implanted seed. Considering the seed shape as a prior, we alleviate the need for a posteriori filtering among candidate shapes. Robustness to noise is enforced by constraining the model to rigid body motion and by privileging image intensity over higher order information. We present encouraging preliminary results on noisy synthetic images. More advanced validation on physical phantoms and clinical images is ongoing.