How To Detect Human Fall In Video

n i l l a f n a m u h t c e t e How to d w e i v r e v o n A ? o e vid Jonas Van den Bergh Toon Goedemé

Jared Willems Mieke Deschodt Koen Milisen Eddy Dejaeger Glen Debard Bert Bonroy Bart Vanrumste

www.fallcam.be www.company.com

Summary • Master project • System overview • Background subtraction • Fall detection • Results • Conclusion & Demo 2

www.fallcam.be

Project Master project “Camera system for elderly fall detection” Part of the TETRA-project: Fallcam www.fallcam.be Goal → Developing a video-based algorithm for fall detection, using grayscale video sequences

3

www.fallcam.be

Basic fall detection system

4

www.fallcam.be

1. Background subtraction

5

www.fallcam.be

Background subtraction

6

www.fallcam.be

Background subtraction Non-recursive techniques Frame differencing Median filtering

Recursive techniques Running average Approximated median filtering Mixture of gaussians

7

www.fallcam.be

AMF & MF Approximated median filtering:

B  x , y=B  x , y1 if I  x , y B  x , y B  x , y=B  x , y−1 if I  x , y ≤B  x , y Median filtering:

Median 8

www.fallcam.be

Calculated Background

Background subtraction

9

www.fallcam.be

Background subtraction In our project median filtering and approximated median filtering → comparative study Median filtering gives better results but is slower than approximated median filtering (0.097 vs 0.20 seconds per frame [intel core 2 duo, T9400]) (MF: 0.35, AMF: 2.44 avg pixel difference)

Buffer of 40 frames, one frame added each 5 frames

10

www.fallcam.be

11

www.fallcam.be

2. Fall detection Three major subdivisions: Extracting parameters from video data Self learning algorithms (Hidden Markov Models) Detection of “abnormal behavior”

We will focus on parameter-based methods

12

www.fallcam.be

Fall detection Aspect ratio Fall angle Vertical projection histograms Centroid of the falling person Horizontal and vertical gradient

13

www.fallcam.be

Fall detection Aspect ratio

14

www.fallcam.be

Fall detection Fall angle

15

www.fallcam.be

Fall detection Vertical projection histograms

16

www.fallcam.be

Further processing Further processing is needed to become a good working fall detection system, possibilities are: Noise suppression (filtering) Morphological operations (erosion / dilation) Shadow detection Ghost detection Occlusion handling Ellipse fitting

17

www.fallcam.be

Results Accuracy: Sideview Frontview

Correct 85% 78%

FP 0% 11%

FN 15% 11%

Tested with 23 sequences

processing speed [s/frame]: AMF 0,36

MF 0,43

Intel Core 2 Duo, T9400, 2,53 Ghz dual core processor

18

www.fallcam.be

Conclusion During this master project, we have developed a working fall detection algorithm Note that all processing was done using grayscale video sequences Better results with additional processing steps Other programming language and more efficient code to speed-up the algorithm

19

www.fallcam.be

Demo

20

www.fallcam.be

Questions?

21

www.fallcam.be