PPM: A Boolean Optimizer for Data Association in Multi-View Pedestrian Detection

Rui Qiu,Ming Xu,Yuyao Yan,Jeremy S. Smith,Yuchen Ling

DOI: https://doi.org/10.1016/j.patcog.2024.110807

IF: 8

2024-01-01

Pattern Recognition

Abstract:To accurately localize occluded people in a crowd is a challenging problem in video surveillance. Existing end-to-end deep multi-camera detectors rely heavily on pre-training with the same multiview datasets used for testing, which compromises their real-world applications. An alternative approach presented here is to project the torso lines of the instance segmentation masks from multiple views to the ground plane and propose pedestrian candidates at the intersection points. The candidate selection process is, for the first time, formulated as a logic minimization problem in Boolean algebra. A probabilistic Petrick’s method (PPM) is proposed to seek the minimum number of candidates to account for all the foreground masks while maximizing the joint occupancy likelihoods in multiple views. Experiments on benchmark video datasets have demonstrated the much improved performance of this approach in comparison with the benchmark deep or non-deep algorithms for multiview pedestrian detection.