We present a novel multi-object tracking algorithm based on multiple hypotheses about the trajectories of the objects. Our work is inspired by Reid?s multiple hypothesis tracking algorithm which is an optimal solution to the motion correspondence that occurs in multi-object tracking. Unfortunately, the exponential growth of the hypotheses tree precludes practical applications. To restrict this growth, many approximations relying on a series of clustering and pruning operations have been proposed. The decisions for these operations are based solely on previous observations and are not guided by observations in later frames. We show that due to multiple splits and merges, relying solely on previous observations to guide these operations may inadvertently eliminate the correct hypothesis. Consequently, this leads to poor tracking performance. To overcome this problem, we determine the validity of a hypothesis by exploiting information in later frames and relating them to previous observations. Experimental results demonstrate the robustness and efficiency of our approach.