Pseudo-label growth dictionary pair learning for crowd counting

摘要

Crowd counting has received increasing attention in the field of video surveillance and urban security system. However, many previous models are prone to poor generalization capability to unknown samples when limited labeled samples are available. To improve or mitigate the above weakness, we develop a novel Pseudo-label Growth Dictionary Pair Learning (PG-DPL) method for crowd counting. To be exact, we treat crowd counting as a task of classification and leverage dictionary learning-based (DL) strategy to target the task. Considering that being short of diverse training samples and imbalanced distribution across different classes in crowd scene inevitably result in large prediction deviation caused by the DL model, we propose to apply pseudo-label growth (PG) and adaptive dictionary size (ADS) to improve the accuracy of crowd counting with limited labeled samples. In the proposed method, PG optimizes the initial prediction via reconstructing the discriminant term to improve the robustness of learned dictionary, while ADS explores the imbalanced distribution among different classes to adapt to the size of class-specific dictionary. Extensive validation experiments on five benchmark databases indicate that the proposed PG-DPL can achieve compelling performance compared to other state-of-the-art methods.