Detecting Kangaroos
in the Wild: The First Step towards Automated Animal Surveillance
Recent studies in computer vision
have provided new solutions to real-world problems in field works. In this
paper, we focus on using computer vision methods to assist with the study of
kangaroos in the wild. Yet there is no kangaroo related dataset for researchers
of interest to use. So we proposed a kangaroo dataset for the kangaroo related
research. Unlike most object detection tasks, kangaroo detection in the wild
possesses markedly more difficult challenges due to the complex background,
varying illumination, occlusion and the multiple poses and extremely deformable
body.
Although kangaroos could drastically
twist their body, we observed that the state-of-the-art DPM method can handle
the challenges. DPM is a star model consists of a root filter and a set of part
filters. The root filter is a coarse representation of the object and the part
filter is to capture the parts of the objects. The parts can have certain level
of deformation around the root filter. DPM can fail to detect kangaroos due to
some parts are missing caused by occlusion and poses, yet the score of these
kind of images are still higher than the true negatives (i.e. the background).
We shall call these as weak negatives (denoted which also include false positives). In DPM, images from will result in a high
FAR.
We propose to use pose-specific SVMs to re-score the images
in. The key idea is to train a pose-specific SVM to deal with
the occluded pattern caused by a certain pose. The multi-pose SVMs serve as an
additional component in the DPM framework. After training of DPM and multi-pose
SVMs (separately), we combine the score from DPM and the multi-pose SVM as
follows in testing process.
Where is our final detector
score. is the highest SVM score among eight pose SVMs (front, rear,
left, right, front-left, front-right, rear-left and rear-right). We further
define two threshold T1>T2 from cross validation. Only images whose DPM
score falls between T1 and T2 will be put into.
As we can see from the table, the proposed approach achieves
marked improvement over standard DPM. The challenge posed by the dataset is
also reflected in the performance of HOG+SVM method.
Tab.1 Detection result on Kangaroos.
In the second evaluation, we used our proposed approach in a Kangaroo
population study. We collected 4,000 frames from four different locations to
get the accumulated kangaroo population distribution over the 24 hours. From
this result, we found that kangaroos tend to be more active in the early
evening in these locations and is much more active in the night than day time.
This finding corroborates previous biological studies suggesting that most
kangaroos are nocturnal animals.
Fig.1 Kangaroo activity pattern
Download
Our data were collected in several
national parks across Queensland State during 2013. In the initial stage, 1,900
cropped images are selected from the raw frames which include other animals
such as dingo, wild cat and emu. You can download the dataset from here: kangaroo_dataset.zip.