This paper describes a innovative algorithm for the on-line estimation of the image jacobian based on previous movements. It has been utilized for the tracking of an uncalibrated 3 DOF joint system, using two weakly calibrated fixed cameras. The algorithm incorporates the epipolar restriction of the two cameras conveyed in the fundamental matrix, and a reliability factor has been defined for the previous movements. The tests carried out show that the proposed algorithm is more robust in presence of noise than those obtained from already existing algorithms in specialized literature.