Towards Mitigating Visual Ambiguity in Person Re-Identification

Abstract

One of the common task in any video surveillance system is re-identifications of humans, which aims at recognizing individuals across disjoint camera views that are spanned over a larger geographical area. It is also known, by the scientific community, as “Person Re-Identification’’ or “People Re-Identification’’ or “Pedestrian Re-Identification’’. In this thesis, we term it as Person Re-Identification (PRId).

In last one decade, PRId has become an intriguing topic of research and has engrossed the computer vision community completely for the severity of challenges it involves like arbitrary pose changes, illumination variations, viewpoint variations, partial occlusions, etc. In this thesis, frameworks on PRId are suggested to mitigate the visual ambiguity arising out of the above challenges.

A person re-identification framework is proposed to represent a pedestrian by some discriminate signature which is formulated by fusing the symmetry based local features. A spectral matching technique satisfying the pairwise constraints is further employed to establish the correspondence between these signatures. This framework provides invariance against the pose changes, light color transformation, illumination and viewpoint variations.

A neuromorphic person re-identification (NPRId) system is developed which concentrates on three aspects such as discovering a set of complementary cues to strengthen the feature descriptor, discovering multiple consensus clusters for the gallery set with high intra-cluster and low inter-cluster similarity, and identifying a subset of gallery candidates with the high probability of retrieving true match. NPRId handles visual ambiguity in similar appearance by re-identifying a target person with in the look-alike gallerycandidates.

A multi-shot person re-identification framework is developed based on the principle of NPRId system which concentrates on the utilization of orientation and appearance cues of pedestrian images to handle the viewpoint challenges. For a given probe, an inlier-intersection set of gallery candidates is generated which holds both the orientations and appearance characteristics similar to that of the probe and reduces the search space. This framework jointly handles issues related to orientation and appearance.

Another multi-shot re-identification framework is proposed to address the illumination changes and scale variations in the appearances of pedestrian image sequence and yields an effective feature representation. Subsequently, a hypergraph structure reflecting the high order relationship between probe and gallery is optimized to learn the similarity scores.

The performances of proposed schemes are experimentally validated on standard data sets and exhibit superior recognition rates when compared with their respective counterparts.