Improving Helmet-Wearing Detection with Human Detection
Improving Helmet-Wearing Detection with Human Detection
复制标题
DOI：
10.1109/nicoint59725.2023.00012
复制
发表时间：
2023-06
期刊：
影响因子：
--
通讯作者：
Chao Zhang;Hiroshi Kawashima;Junzhou Yu;Chunzhi Gu
中科院分区：
文献类型：
--
作者：
Chao Zhang;Hiroshi Kawashima;Junzhou Yu;Chunzhi Gu
戴头盔检测任务需要从给定的图像或视频中检测一个人是否戴着头盔。现有的深度学习研究都存在一个问题，即当目标人的分辨率变低时，检测性能会下降。此外，提高性能还需要神经网络模型的训练成本和数据收集的人工成本。为此，我们提出在不增加训练成本的情况下，使用预先训练好的现成人体检测模型来提高头盔检测的性能，该方法简单而有效。具体来说，利用人体检测结果和头盔佩戴检测结果之间的位置关系对头盔进行重新识别，该位置关系是基于观察到头盔应该在人的边界框内。在实验中，我们证实，特别是在低分辨率下，我们的方法可以显著提高模型的召回率，并进一步提高F1分数。
The helmet-wearing detection task requires detecting whether a person is wearing a helmet or not from a given image or video. Existing studies using deep learning share the problem that the detection performance degrades when the resolution of the target person becomes low. In addition, the training cost of neural network models and the labor cost of data collection are required to improve the performance. To this end, we propose to improve the performance of helmet-wearing detection using a pre-trained off-the-shelf human detection model without additional training cost, which is simple yet effective. Specifically, the helmet is re-identified using the positional relationship between the results of human detection and helmet-wearing detection, which is based on the observation that a helmet should be within the bounding box of a person. In the experiment, we confirm that, especially at low resolution, our method can significantly improve the recall of the model and further improve the F1 score.