Dynamic object construction using belief function theory

A novel approach for multi-object tracking using evidential representation for objects

“…Finally, most sophisticated geometric representations deal with articulated objects (set of sub parts collected by junctions) as sets of geometric shapes (cylinder or ellipse for each sub part) or skeletons [1]. This work extends one of our previous studies namely [35] where an evidential approach has been proposed to construct an object from fragmentary detections (i.e. subsets of pixels that are assumed to be a subpart of an object).…”

“…However, object fragments are non-dense, as opposed/contrarily to superpixels. Besides, in [35], due to computation constraints, fragments are rectangular windows (2D-tiles), thus not respecting the boundary adherence property. Then, from detection fragments that are uncertain since they may correspond either to false positives/false detections that we denote as false alarms in the rest of the paper or to actual subparts of the objects of interest, objects are reconstructed taking advantage of the accumulation through time.…”

“…To manage imprecise and uncertain data (objects characteristics, set of objects, false alarms...), belief function theory is a suitable framework [39]. It has been effective in developing solutions for several problems such as data association and filtering [34], [35]. To reduce the uncertainty of objects, an evidential filtering approach has been proposed in [34].…”

“…Then, we adopt a sparse representation of Ω: instead of enumerating all its possible elements, we only focus on the focal elements of the considered bba. Following the works [3], [35], we represent (non-unequivocally) a focal element by a subset of rectangular subsets of pixels (windows or boxes). However, with such a representation, the bit-wise operators cannot be used (as usually when handling the binary representation of all Ω elements) to derive the basic set operators (inclusion, intersection, union) that have thus to be redefined in an efficient way.…”

“…In order to improve the relevance of the appearance model, the used measure of distance (on which data association will be based) between two objects/people observed at different instants must be robust to partial occlusion and to background changes. The contribution of this paper consists in extending the work of [35] to construct tracklets robust to object crossing (i.e., limited-time occlusions). We use a distance robust to missing subparts of the objects and we take into account the local uncertainty of the object subparts.…”