An Effective Pipeline for Pedestrian Detection in Mid-High Density Crowd
Yu Hao, Duke Gledhill, Ying Liu, Jiulun Fan, Zhijie Xu · 2019
A novel framework is introduced to handle the pedestrian detection in mid-high crowd density. This framework exploits the samples obtained with the unsupervised approach, extracts the combined pattern of Histograms of Oriented Gradients (HOG) and Local Binary Pattern (LBP) for the training and detection. The motion information is utilized to adjust the confidence score of detect annotation, and the soft-NMS is used to handle the balance between the removing of redundant annotation and occlusion. The experiments indicate the proposed approach achieved a promising result compared to state-of-the-art trained with the benchmarking dataset.