When humans predict the arrival time of an object at a designated position in space (i.e. estimate its time-to-collision, TTC), certain biases and characteristics are observed consistently. First, when two objects approaching at different constant speeds have an identical actual TTC, the TTC for the faster object is estimated to be longer than that for the slower object (speed/distance effect). Second, the intraindividual variability of the estimated TTCs increases with the mean estimated TTC (scalar property). Third, short TTCs tend to be overestimated and longer TTCs underestimated (central tendency). Here, we propose a simple Bayesian observer model in which the perceived distance/speed is assumed to be a weighted average of the distance/speed presented in the current trial and a prior determined by the distribution of distances/speeds presented in the experimental session, all processed on a logarithmic scale, with additive Gaussian internal noise. The estimated TTC is set to the ratio of the perceived distance to the perceived speed. We present simulation results and model fits to data from an experiment measuring visual TTC estimation for approaching vehicles in a virtual environment. The results show that both a speed and a distance prior are needed to predict the patterns reported in the literature and observed in the experiment. For the experimental data, the model captures both "molar" (mean estimated TTCs and their intraindividual variability) and "molecular" (cue weights estimated with a reverse-correlation approach) psychophysical results. We conclude that this type of model is a promising approach for predicting human time-to-collision estimation.