Object tracking and identification in video streams with snakes and points

This paper presents a generic approach for object tracking and identification in video sequences, called SAP. The object is described with two image primitives: first, its content is described with Points of interest that are automatically extracted and characterized according to an appearance-based model. Second, the object's envelope is described with a Snake. The originality of SAP consists in a complementary use of these primitives: the snake allows to reduce the points extraction to a limited area, and the point description is efficiently exploited during the snake tracking. Such a characterization is robust to wide occlusions and can be use for object identification and localization purposes. SAP has been implemented with the aim of achieving near real-time performance.