Pedestrian detection task is generally composed of location and classification. In this work, we consider the problem of pedestrian detection under the condition of blurred pedestrian targets and small pedestrian scales. Most methods are only suitable for the detection task under ideal conditions or simply consider one of the above problems. For the sake of achieving better results, we proposed a new method that improved from both location and classification aspects. First, we propose a video-based pedestrian detection method that generates pedestrian proposals from the static feature extraction branch and motion information extraction branch. Secondly, a two-scale classification method is used in the structure to solve the various scale instance problem. Finally, experiments on public datasets demonstrate that this work delivers better performance than the HOG+SVM, SAF R-CNN, and YOLO+GMM methods. © 2022 Institute of Electrical Engineers of Japan. Published by Wiley Periodicals LLC.

中文翻译：

---

一种小而模糊目标的双分支行人检测方法

行人检测任务一般由定位和分类两部分组成。在这项工作中，我们考虑了行人目标模糊和行人尺度小的情况下的行人检测问题。大多数方法只适用于理想条件下的检测任务或简单地考虑上述问题之一。为了获得更好的结果，我们提出了一种从定位和分类两个方面进行改进的新方法。首先，我们提出了一种基于视频的行人检测方法，该方法从静态特征提取分支和运动信息提取分支生成行人建议。其次，结构中采用了双尺度分类方法来解决多尺度实例问题。最后，在公共数据集上的实验表明，这项工作比 HOG+SVM、SAF R-CNN 和 YOLO+GMM 方法提供更好的性能。© 2022 日本电气工程师协会。由 Wiley Periodicals LLC 出版。