Snakes and Spiders

Intensity information is a strong cue for segmentation but on its own cannot be used to distinguish between accidental and non-accidental alignments in a scene, thus resulting in incorrect segmentations. However, motion information can be used to distinguish between accidental and nonaccidental alignments. In this paper an integrated method using both intensity and motion information for the segmentation and tracking of objects in a sequence is presented. The method is based on an extension to active contours (snakes) called spiders. This paper deals with the problems of motion tracking and object segmentation in an integrated common framework. The techniques presented here are based on the observation that segmentation is easier if features have already been successfully tracked over several frames, and tracking is easier if segmentation has already been performed. This suggests an integrated approach to both problems. Feature points on a single rigid object are often connected by quite strong edges and this can be used as a useful cue for segmentation. However, accidental alignment may also cause feature points on separate objects to be connected by a strong edge, so any segmentation using edge strength between feature points will not be able to discriminate between real connections and accidental ones. On the other hand, tracking of moving objects over an image sequence will eventually lead to any accidental alignments becoming non-aligned. Therefore, it seems reasonable to expect that reliable segmentation (and therefore tracking) is best achieved using a combination of intensity and motion cues. It is this conjecture which is addressed in this paper. Motion based segmentation techniques typically come in two flavours: region based methods (usually based on optical flow); and feature based methods. This paper describes a system of the latter type. Combined techniques are also possible and Paragios and Deriche [6] describe a promising technique using an extension of geodesic active contours for segmentation and tracking in a video surveillance application. Although it appears to work very well, their method relies on relatively static backgrounds common in surveillance applications. In contrast, the technique presented in this paper does not rely on static backgrounds and is therefore likely to be more useful for mobile robotic applications. Smith and Brady [8] describe ASSET-2, a real time motion segmenter and tracker. ASSET-2 utilises corner points which are matched between frames and clustered to produce an object segmentation. The object clusters are then used to improve the reliability of tracking in future frames. The limitation of this system is that until reliable clusters are formed, each feature must be tracked individually. Of course, it is more difficult to track individual features robustly. The system described in this paper does not suffer from this limitation since features are initially tracked dependent on their neighbours according to the intensity information linking any two of them. In other words, corner points linked by strong edges will attempt to remain at a similar distance in subsequent frames.