EDF and OCCU Fall Detection Datasets

Abstract

Falls are one of the major risks for seniors living alone at home. Fall detection has been widely studied in the computer vision community, especially since the advent of affordable depth sensing technology like the Kinect. Most existing methods assume that the whole fall process is visible to the camera. This is not always the case, however, since the end of the fall can be completely occluded by a certain object, like a bed. For a system to be usable in real life, the occlusion problem must be addressed. To quantify the challenges and assess performance in this topic, we present an occluded fall detection benchmark dataset containing 60 occluded falls for which the end of the fall is completely occluded.

Data Collection

EDF is a non-occlusion fall detection dataset, and OCCU is an occlusion fall detection dataset that were created using Kinect cameras for XBOX 360 with the Microsoft Kinect for Windows SDK Beta at a frame rate of about 30fps. Each of the 5 subjects performed a non-occluded fall along eight directions in each viewpoint in the EDF dataset. Each of the 5 subjects performed six occluded falls in each viewpoint in the OCCU dataset. In the OCCU dataset, the end of the fall is completely occluded by a bed. 

EDF

The EDF dataset is comprised of 25,881 frames and 40 real falls in videos from the first viewpoint, and 24,497 frames and 40 real falls in videos from the second viewpoint. The two viewpoints were recorded at the same time, and thus every event was recorded simultaneously from both viewpoints. Our subjects also performed a total of 30 actions that tend to produce features similar to those of a fall event, namely: 10 examples of picking up something from the floor, 10 cases of sitting on the floor, and 10 examples of lying down on the floor.

OCCU

The OCCU dataset includes 25,618 frames and 30 totally occluded falls in videos from the first viewpoint, and 23,703 frames and 30 totally occluded falls in videos from the second viewpoint, performed by the same subjects. Each viewpoint was recorded at separate times from the other viewpoint, and thus we had no instances where the same events were recorded simultaneously from both viewpoints. Our subjects also performed a total of 80 actions that tended to produce features similar to those of a fall event, namely: 20 examples of picking up something from the floor (all of them are non-occluded), 20 examples of sitting on the floor (all of them are non-occluded), 20 examples of tying shoelaces (all of them are non-occluded), and 21 examples of lying down on the floor(all of them are totally occluded at the end frame).

If you use EDF and/or OCCU in your work, please cite the original publication that presents these datasets:

Zhang, Z., Conly, C., Athitsos, V. (2014). Evaluating Depth-Based Computer Vision Methods for Fall Detection under Occlusions. In: Bebis, G., et al. Advances in Visual Computing. ISVC 2014. Lecture Notes in Computer Science, vol 8888. Springer, Cham. https://doi.org/10.1007/978-3-319-14364-4_19