Despite the recent advances in multiple object tracking (MOT), achieved by\njoint detection and tracking, dealing with long occlusions remains a challenge.\nThis is due to the fact that such techniques tend to ignore the long-term\nmotion information. In this paper, we introduce a probabilistic autoregressive\nmotion model to score tracklet proposals by directly measuring their\nlikelihood. This is achieved by training our model to learn the underlying\ndistribution of natural tracklets. As such, our model allows us not only to\nassign new detections to existing tracklets, but also to inpaint a tracklet\nwhen an object has been lost for a long time, e.g., due to occlusion, by\nsampling tracklets so as to fill the gap caused by misdetections. Our\nexperiments demonstrate the superiority of our approach at tracking objects in\nchallenging sequences; it outperforms the state of the art in most standard MOT\nmetrics on multiple MOT benchmark datasets, including MOT16, MOT17, and MOT20.\n