Detection Of Moving Objects From Moving Platform

The detection and tracking of moving objects is vitally important for both military and civil application, and consequently studying how to highly-automatically detect and track them is very indispensable. Here first we summarize now-existent detection and tracking algorithms, and then analyze their disadvantages and applicable environments, combining with current research requirements finally present one new algorithm based on optical flow technology. Considering moving platform and shortcomings of optical flow computation, compensating background motion by means of image registration is necessary to weaken its adverse influence as largely as possible. Here two mainly-focused problems about image registration are that molding background motion as affine transformation more consistent with actual application or by far simple pure translation is better, and meanwhile that single template or multiple templates are preferable. It is well known that computing optical flow is very time-consuming, and thus which optical flow computation algorithm should be adopted is very important. Generally speaking, the higher exact algorithm requires the more time for complex computation. Of course high exact optical flow is very interesting, whereas too low efficiency is also unadvisable, and thus classical HS and improved algorithms are carefully discussed, where first analyzing two constraints including basic equation constraint and smoothness constraint, then deducing computation formula, and last changing basic equation into conditional constraint, and thus get one finally-adopted iterative formula. The last part is detecting moving objects using above optical flow, owing to incomplete background compensation and shortcomings of optical flow computation, firstly necessarily filtering optical flow to clear noises away, after that segmenting objects by thresholding, later on proposing one brand-new centroid algorithm for later rapidly finding out all possible objects, and then merging them because of the incontinuity brought by object segmentation, at last detecting real moving objects for the fact that noises are random.