Abstract This thesis deals with automatic localization of thin surgical tools such as needles or electrodes in 3D ultrasound images. The precise and reliable localization is important for medical interventions such as needle biopsy or electrode insertion into tissue. The reader is introduced to basics of medical ultrasound (US) imaging. The state of the art localization methods are reviewed in the work. Many methods such as Hough transform (HT) or Parallel Integral Projection (PIP) are based on projections. As the existing PIP implementations are relatively slow, we suggest an acceleration by using a multiresolution approach. We propose to use model fitting approach which uses randomized sample consensus (RANSAC) and local optimization. It is a fast method suitable for real-time use and it is robust with respect to the presence of other high-intensity structures in the background. We propose two new shape and appearance models of tool in 3D US images. Tool localization can be improved by exploiting its tubularity. We propose a tool model which uses line filtering and we incorporated it into the model fitting scheme. The robustness of such localization algorithm is improved at the expense of additional time for pre-processing. The real-time localization using the shape model is demonstrated by implementation on the 3D US scanner Ultrasonix RP. All proposed methods were tested on simulated data, phantom US data (a replacement for a tissue) and real tissue US data of breast with biopsy needle. The proposed methods had comparable accuracy and the lower number of failures than the state of the art projection based methods.