Anchor-based group detection in crowd scenes

Patch-based topic model for group detection

“…In this part, we evaluate the performance the proposed PTM on group number estimation. Two standard metrics, averaging difference (AD) and variance (VAR) [11] are used as measurements. A lower value of AD indicates the less deviation from the ground truth, and a lower VAR means a higher stability on the group number estimation.…”

“…To analyze the crowd behavior, it is fundamental to extract the individuals in crowd scenes. Since detection and tracking algorithms are inapplicable in crowd scenes, previous studies mostly treat particles [4,5] or feature points [6][7][8][9][10][11] as study objects, and model their velocities directly to detect groups. But both the particles and feature points are too microcosmic to reflect the global crowd motion.…”

“…Thus, the investigation of the underlying prior could facilitate group detection. However, many existing studies [4][5][6][7][8]11] only emphasize the observed movements of individuals, and neglect the underlying prior.…”

A new approach for social group detection based on spatio-temporal interpersonal distance measurement

“…One category clusters pedestrians based on different motion patterns of crowds [27] , [28] , [29] , [30] , [31] , [32] , with wide applications in medium and high-density crowd scenes, such as crowd flow monitoring and crowd behaviour analysis. For example, Shao et al [27] adopted a robust group detection algorithm and a rich set of group-property visual descriptors through learning the collective transition prior; they then utilised visual descriptors to quantify group-level properties for crowd understanding [28] .…”

“…For example, Shao et al [27] adopted a robust group detection algorithm and a rich set of group-property visual descriptors through learning the collective transition prior; they then utilised visual descriptors to quantify group-level properties for crowd understanding [28] . Chen et al [29] proposed an anchor-based manifold ranking (AMR) method to classify individuals into local clusters according to topological relationship to the anchors, and exploited a coherent merging strategy to recognise global consistency in crowed scenes. Wang et al [30] designed a multi-view clustering method for group detection by combining the orientation and context similarities of feature points, whereas Han et al [31] developed a crowd activity discovery algorithm to explore latent action patterns among crowd activities and clustering them.…”