Nature Conservation Drones

Nature Conservation Drones for Automatic Localization and

Counting of AnimalsCamiel R. Verschoor - @Camiel_Vvan Gemert, Verschoor, Mettes, Epema, Koh & Wich

6 September 2014

Motivation

Accurate monitoring key ingredient of nature conservation

Animal monitoring involves:

• Animal counting

• Indirect counting of animal signs

Conventional ground surveys can be time consuming

Aerial surveys expensive, not available or unpractical

2

Conservation Drones

Unmanned Aerial Vehicles are cheap, accessible and autonomous (Jones IV, Pearlstine, Percival, 2006).

Generates numerous photos and videos

3

Computer Vision

Strong need for automated detection of objects.

• Relatively small

• Skewed vantage point

!

This paper evaluated how current object detection methods scale to drone nature conservation tasks

4

Related work

Conservation Drones

• Conservation (Koh & Wich, 2012)

• Terrain mapping (Getzin, Wiegand & Schning, 2012; Hodgson, Kelly & Peel, 2013)

Computer Vision

• Convolutional networks (computational intensive) (Girshick, Donahue, Darrell & Malik, 2014; Sermanet et al., 2014)

• Bag of Words/Fisher Vectors (memory intensive) (Uijlings, van de Sande, Gevers & Smeulders, 2013; Cinbis, Verbeek & Schmid, 2013)

5

Evaluating Nature Conservation

Two tasks: Animal Detection and Animal Counting.

Recorded dataset:

• Pelican with GoPro

• Two separate flights

• 4 training videos (12,673 frames) 2 test videos (5,683 frames)30 unique animals

• Manually annotated with vatic. (Vondrick, Patterson & Ramanan, 2012)

6

Methods

Object detection

• Deformable part-based model (DPM)DPM & Color DPM (Felzenszwalb, Girshick, McAllester & Ramanan, 2010; Khan et al., 2012)

• Exemplar SVM (Malisiewicz, Gupta & Efros, 2011)

Animal Counting

• KLT Tracker (Lucas, Kanade et al., 1981)

• Merge detections when: (Everingham, Sivic & Zisserman, 2009)

7

Results - Proposal Quality

• Selective search (Uijlings, van de Sande, Gevers & Smeulders, 2013)

• Search time is not significantly decreased

• Object proposal-based detection systems are from a computational standpoint not suitable

8

Results - Animal Detection

9

Results - Animal Detection

Exemplar SVM DPM Color DPM10

Results - Animal Counting Ground Truth

11

Conclusion

• Investigated automatic object detection methods for nature conservation

• Three lightweight detection methods are benchmarked for animal detection giving promising results

• Results show that animal counting is a difficult task

• Nature Conservation Drones have potential!

12

Questions

Remember! Anything can Fly!

Camiel R. Verschoor - @Camiel_V [email protected]