In this paper we discuss the feasibility of using a DCNN to implement Fall and Head detection for a Danger Management System. For this propose, AlexNet and Inception - v3 models in MatConvNet and TensorFlow libraries were used for training new DCNNs with Fine - Tuning method. Additionally, a new public dataset was created, which includes diverse fall poses, as well as, top views of people walking in a scene.
The presented dataset has a total of 155,530 images. These images were obtained through the recording of members of CIDIS, in 4 sessions. In total, 10 videos with a duration of 4 minutes each were obtained. The participants were asked to bring different clothes, in order to give variety to the images. After this, the frames of the videos were separated at a rate of 5 frames per second. All these images were captured from a top view perspective. The original images have a resolution of 1280x720 pixels. Subsequently, a classification was made to decide which images could be used for the training. All these images were captured from a top view perspective.
For more information about the dataset, go to Readme.txt
|
|
| Download Dataset |
@article{, title = "An Empirical Comparison of DCNN libraries to implement the Vision Module of a Danger Management System", journal = "International Conference on Deep Learning Technologies (ICDLT 2017)", pages = "60-65", year = "2017", isbn = "978-1-4503-5232-1", doi = "10.1145/3094243.3094255", url = "http://refbase.cidis.espol.edu.ec/files/siannapuente/2017/56_SiannaPuente_etal2017.pdf", author = "Sianna Puente, Cindy Madrid, Miguel Realpe, Boris X. Vintimilla", keywords = "DCNN, Fall detection, Head detection, Tensorflow, Matconvnet, Danger Management Systems", }