Automated Fall Detection on PrivacyEnhanced Video Alex Edgcomb
Automated Fall Detection on Privacy-Enhanced Video Alex Edgcomb Frank Vahid University of California, Riverside Department of Computer Science Copyright © 2012 Alex Edgcomb, UC Riverside. 1 of 12
Reasons to detect falls with privacyenhanced video Body-worn +Anywhere -Not always worn Detect other events Copyright © 2012 Alex Edgcomb, UC Riverside. Privacy adjustable 2 of 12
Efficient person-detection in video via foreground-background segmentation Background image = Video frame Copyright © 2012 Alex Edgcomb, UC Riverside. Foreground 3 of 12
Abstracting person to rectangle Video frame Foreground Minimum bounding rectangle (MBR) of foreground Copyright © 2012 Alex Edgcomb, UC Riverside. 4 of 12
Fall shown as four MBR features Copyright © 2012 Alex Edgcomb, UC Riverside. 5 of 12
Fall classification (details in paper) Dynamic time warping Observed shape Similarity 0. 84 Binary tree classification Observed shape 0. 46 Characteristic fall shape Non-fall 0. 88 DTW established time series technique Non-fall Copyright © 2012 Alex Edgcomb, UC Riverside. Fall 6 of 12
Recordings gathered • • 23 recordings (12 fall, 11 non-fall) Sole male twenty-six year old actor Recorded in living room Recorded with webcam @ 15 fps Copyright © 2012 Alex Edgcomb, UC Riverside. 7 of 12
Fall detection accuracy by feature Feature Average sensitivity Average specificity Height of MBR in pixels 0. 31 0. 30 Width of MBR in pixels 0. 91 0. 92 Height-to-width ratio of MBR Width-to-height ratio of MBR 0. 44 0. 50 0. 64 0. 67 For each feature, trained binary classifier using leave-one-videoout, then tested with video left out. Copyright © 2012 Alex Edgcomb, UC Riverside. 8 of 12
Fall detection on privacy-enhanced video Raw Blur Silhouette Bounding -box -oval Copyright © 2012 Alex Edgcomb, UC Riverside. 9 of 12
Fall detection accuracy by privacy enhancement Privacy setting Average sensitivity Average specificity Raw 0. 91 0. 92 Blur 1. 00 0. 67 Silhouette 0. 91 0. 75 Bounding-oval 0. 91 0. 92 Bounding-box 0. 82 0. 92 • Auto-converted 23 raw videos into each privacy enhancement • Used trained binary classifier from raw video. • Tested with each privacy enhancement. Copyright © 2012 Alex Edgcomb, UC Riverside. 10 of 12
Characteristic fall shape is nearly identical for raw and privacy-enhanced video Copyright © 2012 Alex Edgcomb, UC Riverside. 11 of 12
Conclusions • Bounding-oval yielded same accuracy as raw video • Privacy-enhanced fall detection is viable • Future work – Compare our algorithm to previous works – Experiments with more recordings – Consider more privacy enhancements Copyright © 2012 Alex Edgcomb, UC Riverside. 12 of 12
- Slides: 12