Data collection
Embedded Files
We provide a novel, comprehensive Vision-based Fallen Person (VFP) dataset to for detecting fallen persons using machine learning algorithms. For dataset generation, we collected 294,713 frames from 178 videos.
Data Annotation Rule
In order to provide the accurate annotations as much as possible including overlap and occlusion cases, we introduce the strict annotation rules.
We assign bounding boxes and labels to the images based on strict and consistent annotation rules regarding different occlusion cases and overlaps.
The rules are enforced to provide the accurate labeling as follows:
Rule 1. Occlusion Labeling Rule.
When the person is occluded but exposes any part of his or her body, we annotate the entire body, referring to Fig.1.(a)
When the overall part of the person is occluded, we annotate the only visible part of the body, referring to Fig.1.(b)
Rule 2. Overlap Labeling Rule.
When two bounding boxes are overlapped and the behind box is in the front box, we do not annotate behind box, referring to Fig.2.(a) and (b)
Figure 1. Example images for occlusion and overlap rules.
De-identification
We de-identify all identifiable car number plates and people shown in our dataset. Basically, as mentioned in main paper section 3.2, the process consist as following: cropping the interested region and applying mosaic followed by blurring. After that, we attach the processed images to the original position.
The results are presented in Fig.2
Figure 2. Example images for de-identification. (a) and (b) are the images after applying de-identification process. (c) is the image that is not applied, as it cannot be recognizable.
Page updated
Google Sites
Report abuse