13. Research Project 1
Nature Conservation Drones for Automatic
Localization and Counting of animals
J.C. van Gemert, C.R. Verschoor, P. Mettes, H.K. Epema, L.P. Koh, and S. Wich.
Evaluates how object detection methods scale to drones
15. Dataset
Data acquisition system
Pelican with GoPro
1080p (1920 x 1080 px)
Data acquisition process
Two separate flights
4 training and 2 test videos
30 unique cows
Dataset challenges
Relatively small objects
Skewed vantage point
Data: www.camielv.nl
16. Detection Methods
Deformable Part-Based Model
(DPM) by Felzenszwalb et al. (2010)
• Object is a composition of
parts
• Proposal score is based on
root and position of the parts
Colour DPM by Khan et al. (2012)
• Adds colour information
17. Detection Methods
Exemplar SVM by Malisiewicz et al.
(2011)
• Trains an SVM model for
every exemplar in training set
• Generalizes and allows
knowledge sharing
All methods use Histogram of
Oriented Gradient features
18. Counting Method
KLT Tracker by Kanade et al. (1981)
• To obtain point tracks over
time of the proposals
Detections are merged using:
by Everingham et al. (2009)
Determines whether detections
belong to the same unique
animal.
22. Research Project 2
Object Detection in Aerial Imagery
A.E.M. Visser, J.C. van Gemert, and C.R. Verschoor (not published)
An object proposal method for aerial imagery
23. Dataset
Data acquisition system
Twinstar with GoPro
7MP (3000x2250 px)
Data acquisition process
Various flights
577 annotated images of
rhinos, zebras, people and
vehicles
Dataset challenges
Small objects
Data: www.dutchuas.nl/dataset
24. Proposal Method
Extract descriptors
• SURF descriptors by Bay et al.
(2006)
• Normalised RGB colour-space
Train SVM Model
• Classifies descriptors into
objects and non-objects.
Density-based clustering
Filters out false positives
• Density-based spatial
clustering of applications with
noise (DB-SCAN)
• Mean shift
27. Results
Pixel reduction: the percentage of pixels not in a bounding box
Selective Search by Uijlings et al. (2013)
Edge Boxes by Zitnick et al. (2014)
28. Future work
Deep Learning on a Conservation Drone
C. Tran, J. van Doorn, J.C. van Gemert and C.R. Verschoor (not published)
Optimising Deep Learning for Nature Conservation