We address the problem of performing 3D motion and structure analysis along long video sequences (mpeg animation, surveillance recording television programs, etc..) in which we want to : - be able to segment the sequence in smaller pieces of image sets corresponding to homogeneous parts of the sequence, - stabilize the sequence in order to reduce the disparity between two consecutive frame, - detect unexpected events such as moving objects or close obstacles, - reconstruct the 3D shape of some predefined objects, which implies the self calibration of the image sequence. Such data have two characteristics : (a) they have been recorded using cameras with unknown and possibly varying optical parameters\footnote{ We are in the un-calibrated case quoted before.} and (b) cameras are moving while some objects are also moving in the scene. It is, in general, not possible to entirely analyze such sequences automatically because of the complexity of the contained information. On the other hand, a complete manual analysis is also not possible because of the huge amount of data. Here, the idea is to pre-process the image sequence, in order to ease its interactive analysis and provide to the user an enhanced image sequence in which: - the image sequence have been stabilized to ease its observation, so that. - first order approximations of usual motion equations could be numerically valid, - disparity not related to object motion or object depth parallax displacement could be eliminated, - regions of disparities corresponding to close obstacles or objects in motion have been segmented,i.e. their contour have been extracted and the relative location/size of these objects have been computed.