TY - GEN
T1 - Evaluation of sampling-based pedestrian detection for crowd counting
AU - Ge, Weina
AU - Collins, Robert T.
PY - 2009
Y1 - 2009
N2 - With increases in computing power, the once computationally expensive sampling-based methods have been successfully applied to solve many hard vision problems of combinatorial nature, such as image segmentation, video tracking, and object detection. In this paper, we perform pedestrian detection using the reversible jump Markov Chain Monte Carlo (RJMCMC) sampling method. A crowd scene is viewed as a realization of a Marked Point Process (MPP) that consists of a random set of people in a bounded region. Each person is associated with a random 'mark' that governs their location and size in the image. To automatically infer the number of people in the scene and their spatial locations, RJMCMC is used to sample person hypotheses from an underlying stochastic process and evaluate them against the image observation to find the optimal configuration that best explains the image. We further extend the detector to hypothesize people not in the image plane but in 3D space, by incorporating multi-view information. The detection performance of both the single- and multi-view versions are evaluated on the crowd counting task in the PETS 2009 dataset.
AB - With increases in computing power, the once computationally expensive sampling-based methods have been successfully applied to solve many hard vision problems of combinatorial nature, such as image segmentation, video tracking, and object detection. In this paper, we perform pedestrian detection using the reversible jump Markov Chain Monte Carlo (RJMCMC) sampling method. A crowd scene is viewed as a realization of a Marked Point Process (MPP) that consists of a random set of people in a bounded region. Each person is associated with a random 'mark' that governs their location and size in the image. To automatically infer the number of people in the scene and their spatial locations, RJMCMC is used to sample person hypotheses from an underlying stochastic process and evaluate them against the image observation to find the optimal configuration that best explains the image. We further extend the detector to hypothesize people not in the image plane but in 3D space, by incorporating multi-view information. The detection performance of both the single- and multi-view versions are evaluated on the crowd counting task in the PETS 2009 dataset.
UR - http://www.scopus.com/inward/record.url?scp=77950599752&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=77950599752&partnerID=8YFLogxK
U2 - 10.1109/PETS-WINTER.2009.5399553
DO - 10.1109/PETS-WINTER.2009.5399553
M3 - Conference contribution
AN - SCOPUS:77950599752
SN - 9781424455034
T3 - Proceedings of the 12th IEEE International Workshop on Performance Evaluation of Tracking and Surveillance, PETS-Winter 2009
BT - Proceedings of the 12th IEEE International Workshop on Performance Evaluation of Tracking and Surveillance, PETS-Winter 2009
T2 - 12th IEEE International Workshop on Performance Evaluation of Tracking and Surveillance, PETS-Winter 2009
Y2 - 7 December 2009 through 9 December 2009
ER -