Crowd pedestrian detection using expectation maximization with weighted local features
Shih-Shinh Huang, Chun-Yuan Chen · 2017
This study proposes a method for crowd pedestrian detection based on monocular vision using expectation maximization (EM) with weighted local features. The proposed method mainly consists of two stages: training and detection stages. During training stage, the proposed method firstly constructs a model for describing the pedestrian appearance based on a set of salient features. During detection stage, an algorithm called expectation maximization (EM) is applied to group the extracted corners to several pedestrians based on the constructed codebook through performing E-step and M-step iteratively. The use of EM algorithm makes the proposed method be capable of detecting partially occluded pedestrians, especially in crowded scenes. In the experiment, a well-known dataset called CAVIAR is used to validate the effectiveness of the proposed method.