Pedestrian Detection in Crowded Scenes Based on Cascade R-CNN
Gaoxiang Yu, Nan Xiang, Chuanzhong Pan · 2022
The occlusion in crowded scenes and the interference of similar objects in the background are one of the main reasons that lead to missed pedestrian detection. In response to this problem, an improved Cascade R-CNN pedestrian detection algorithm using dynamic regressors is proposed. Firstly, the offset of each sample is adjusted to a dynamic offset and normalization to strengthen the regression performance of each regressor. Then, based on the preliminary regression, a secondary detection module of occluded pedestrians is constructed to further enhance the discrimination between pedestrians and pedestrians. The proposed algorithm performed ablation experiments on the datasets Caltech, CityPersons and CrowdHuman, and the missed detection rates were reduced by 21.3%, 5.0% and 8.4% on the Heavy subset of the above three datasets, respectively. The experimental results show that the improved Cascade R-CNN algorithm is strongly robust to obstructed pedestrians.