Many of the video image trackers today use the centroid as the tracking point. In engineering, a target's centroid is computed from a binary image to reduce the processing time. Hence thresholding of gray level image to binary image is a decisive step in centroid tracking. How to choose the feat thresholds in clutter is still an intractability problem unsolved today. This paper introduces a high-accuracy real-time automatic thresholding method for centroid tracker. It works well for variety types of target tracking in clutter. The core of this method is to get the entire information contained in the histogram, such as the number of the peaks, their height, position and other properties in the histogram. Combine with this histogram analysis; we can get several key pairs of peaks which can include the target and the background around it and use the method of Otsu to get intensity thresholds from them. According to the thresholds, we can gain the binary image and get the centroid from it. To track the target, the paper also suggests subjoining an eyeshot-window, just like our eyes focus on a target, we will not miss it unless it is out of our eyeshot, the impression will help us to extract the target in clutter and track it and we will wait its emergence since it has been covered. To obtain the impression, the paper offers a idea comes from the method of Snakes; it give a great help for us to get a glancing size, so that we can compare the size of the object in the current frame with the former. If the change is little, we consider the object has been tracked well. Otherwise, if the change is bigger than usual, we should analyze the inflection in the histogram to find out what happened to the object. In general, what we have to do is turning the analysis into codes for the tracker to determine a feat threshold. The paper will show the steps in detail. The paper also discusses the hardware architecture which can meet the speed requirement.