A person's ability to detect a weak visual target stimulus varies from one viewing to the next. We tested if the trial-to-trial fluctuations of neural population activity in the human brain are related to the fluctuations of behavioral performance in a 'yes-no' visual motion detection task. We recorded neural population activity with whole-head magnetoencephalography (MEG) while subjects searched for a weak coherent motion signal embedded in spatiotemporal noise. We found that, during motion viewing, MEG activity in the 12-24 Hz ('beta') frequency range is higher, on average, before correct behavioral choices than before errors, and that it predicts correct choices on a trial-by-trial basis. This performance-predictive activity is not evident in the pre-stimulus baseline and builds up slowly after stimulus onset. Source reconstruction revealed that the performance-predictive activity is expressed in the posterior parietal and dorsolateral prefrontal cortices, and, less strongly, in the visual motion-sensitive area MT+. 12-24 Hz activity in these key stages of the human dorsal visual pathway is correlated with behavioral choice in both, target present and absent conditions. Importantly, in the absence of the target, 12-24 Hz activity tends to be higher before 'no' choices ('correct rejects') than before 'yes' choices ('false alarms'). It thus predicts the accuracy, and not the content, of subjects' upcoming perceptual reports. We conclude that beta band activity in the human dorsal visual pathway indexes, and potentially controls, the efficiency of neural computations underlying simple perceptual decisions.