Vision based human fall detection

This project investigates how the performance of a vision based human fall detection system can be improved by using a sparse sampling technique and multi-modal input. The proposed solution consists of two stages: a pre-processing stage, which includes sampling from the trimmed videos, modality extraction and frame augmentation as well as a prediction stage, which uses a BNInception CNN model as a feature extractor and a logistic regression model to produce a prediction for a video.

The influence of using different modalities (RGB images, optical flow frames, and RGB difference images) as the input to the system is investigated during the training of the system on a mixture of the FDD and Multicam datasets. The effect of using different transfer learning techniques from the human action recognition domain are studied by evaluating the best systems on the URFD fall dataset. Finally the best system is tested on a subset of the HMDB-51 dataset which is a more realistic dataset with camera movement.

It is shown that using a fine tuned model with a fused RGB and optical flow modality can improve the performance of the system by allowing it to detect fall specific features. However, the proposed system shows limited capability in the problem of fall detection when tested on real data due to the limited nature of existing fall datasets.


Room A 1.03, Robert-Hooke-Str. 1 in Bremen

In der Regel sind die Vorträge Teil von Lehrveranstaltungsreihen der Universität Bremen und nicht frei zugänglich. Bei Interesse wird um Rücksprache mit dem Sekretariat unter sek-ric(at)dfki.de gebeten.

last updated 30.07.2019
to top