In Fall 2005 we deployed a camera system called the "BearCam" to monitor the behaviour of grizzly bears at a remote location near the arctic circle. The system aided biologists in collecting the data for their study on bears' behavioural responses to ecotourists. We developed a camera system for operating in the challenging arctic conditions. We developed a novel "motion shapelet" algorithm for automatically detecting bears in the video captured by this camera system. This algorithm is an extension of the shapelet features, which are mid-level features capturing pieces of shape. Our extension of this technique incorporates motion information and proves effective at automatically detecting the occurrence of bears.

Wildlife managers use a variety of techniques to monitor wildlife around ecotourism sites. The traditional techniques for population-scale data collection, such as mark-recapture or aerial counts, are labour intensive and extremely costly. As an alternative or complement to these methods, the use of camera systems, which collect information largely in the absence of human operators, is increasing in popularity. However, cameras generate large amounts of data, which are typically sorted manually to collect the required data. As computer vision researchers, there is a great opportunity to aid natural scientists by automating parts of the video analysis process.

We developed a camera system for aiding our biologist partners in collecting data. This system monitored a site in the northern Yukon for hundreds of hours. We used a novel algorithm for automatically detecting bears in the video captured by this camera system. The ability to automatically detect segments of video containing bears can drastically reduce the amount of time needed by biologists to analyze the video. Rather than manually sifting through the hundreds of hours of video, much of it devoid of bear activity, they can focus their time only on the relevant portions.

In the videos which we collect, there are a variety of moving objects including waving trees and running water, and animals such as birds and fish. The "motion shapelet" algorithm automatically learns the pattern of image gradients and presence of motion which is indicative of the presence of a bear rather than one of these other objects. The images below show the "motion shapelets" which our algorithm finds, which are indicative of the presence (left) or absence (right) of bears.


The top two images show the image gradient features which occur on bears and off bears respectively. The bottom two images show the motion features which occur on bears and off bears respectively. For example, the bottom images show that a particular pattern of motion, in a rough bear-like blob shape, is indicative of a bear, while extraneous motion such as that of the trees in the background occurs in a broader less-structured pattern.

We applied our bear detection algorithm to video collected in the Yukon in Fall 2005. The video below shows an example of the results, showing the location in each video frame with the highest rank returned by the bear detection algorithm.