Point-based stereo visual odometry systems typically estimate the camera motion by minimizing a cost function of the projection residuals between consecutive frames. Under some mild assumptions, such minimization is equivalent to maximizing the probability of the measured residuals given a certain pose change, for which a suitable model of the error distribution (sensor model) becomes of capital importance in order to obtain accurate results. This paper proposes a robust probabilistic model for projection errors, based on real world data. For that, we argue that projection distances follow Gamma distributions, and hence, the introduction of these models in a probabilistic formulation of the motion estimation process increases both precision and accuracy. Our approach has been validated through a series of experiments with both synthetic and real data, revealing an improvement in accuracy while not increasing the computational burden.