Online Visual Tracking Using Temporally Coherent Part Cluster

Abstract

Recent advances in visual tracking have focused on handling deformations and occlusions using the part-based appearance model. However, it remains a challenge to come up with a reliable target representation using local parts, and hence existing trackers continue to face drifting problems. To deal with this challenge, we propose a robust online model, formulating the tracking task as a problem of identifying Temporally Coherent Part (TCP) clusters. Specifically, we pose the TCP clusters identification task as a dense neighborhoods searching problem using a relational hypergraph in which the relationship among multiple temporal local parts is encoded as the affinity value of a hyperedge connecting them. Such high-order relationships among multiple local parts across the temporal domain make our tracker more robust towards deformations and occlusions. Extensive experiments on various challenging video sequences demonstrate that our TCP-based method performs better than the state-of-the-art methods.

Geometric Hypergraph Learning for Visual Tracking

“…Although our method is related with three previous works including SPT [34], DGT [6] and TCP [22], there are significant differences between our method and them, which are concluded below.…”

Multi-level visual tracking with hierarchical tree structural constraint

“…Xie et al [19] employ a minimum spanning tree to model the geometric structure constraint between superpixels, and infer the target state by voting from confident parts of the target based on appearance similarity and structural constraint. In addition to leveraging the spatial relationships between local parts, Li et al [5] construct a relational hypergraph, which models the high-order relationships among multiple local parts across the temporal domain, and identify temporally coherent parts for the target representation.…”

“…Among the challenging issues, occlusion and deformation have drawn much attention, because they are ubiquitous in real world tracking scenarios and remain unsolved [1]. Recently, an increasing number of tracking algorithms are proposed to solve the occlusion and deformation problems, of which part-based model is mostly adopted by stateof-the-art trackers [1][2][3][4][5][6][7][8][9].…”

“…Different strategies have been proposed to generate local parts. According to the level of granularity, part-based target representation can be categorized into three levels: top level (holistic target models [10][11][12]), middle level (superpixels [1,5,3] and patches [13,2,6,9]), and bottom level (keypoints [7,14,15] and pixels [16,17]). Each level target representation has its pros and cons.…”

Geometric structure-constraint tracking with confident parts

“…[11] exploits multiple levels to quantify appearance of object and finds the most possible position of the target by jointly classifying the pixels and superpixels and obtaining the best configuration across all levels. Compared with sole pixel-level representation, the employment of mid-level representation makes the model applicable for generic object and robust to changes such as object motion, lighting conditions and occlusion in running, so we select mid-level feature as basic representation of object in this paper.…”