Automatic Wildlife Sensing

Author: Stephen Wrathall
Supervisor: Andrew Taylor
Research Theme: Information & Communications

Aim

To develop a robust method to automatically classify thousands of high-resolution images into those which contain animals, and those which do not.

Background & Motivation

Andrew Taylor (CSE) has constructed a solar-powered wildlife monitoring system which captures high-resolution images whenever a scene change is detected. This system can run for days, weeks or months capturing many thousands of images. The problem is that clouds, wind, rain, lightning and flying insects also trigger image capture so most of the captured images do not contain wildlife.

My task was to construct software which automatically processed these high-resolution images using probabilistic, empirical and other researched methods to classify them into animal and non-animal groups.

Left & Right: Shows program detects shade changes instead of animals.

Challenges

How hard is it to detect an animal? Can't you just compare one image containing an animal, with another image without an animal, subtract the two images and you'll see where the animal is? The answer to this is yes, you can. Does it always work? No. This is because usually, animals do not contribute to the largest changes in an image. In fact, the largest and most common changes that occur are due to sunlight and the effects of wind.

For example, when the ground is cast under shadow by a tree, it appears almost black. If the tree blows so the ground is exposed to sunlight, the biggest change possible has occurred from the computer's point of view - from black to white - but there was no animal. So the research is not in detecting animals, it is detecting what changes are not due to animals.

Shade of an animal is detected as an actual animal.		Sunlight causes huge changes, making it hard to detect whether a large change is due to an animal, or the weather.

Method & Results

The three main methods I took were shade detection, cumulative change, and box histograms.

Left & Right: A pelican detected using the box histograms and cumulative change model.

Person detected using box histograms		Person detected using cumulative change model. Trails of her steps in the dirt were detected in both images.

1. Shade Detection

Shade detection involved converting colour spaces from Red Green Blue (RGB) to Hue Saturation Value (HSV). Using the HSV colour scheme, I could calculate a pixel's colour with the H and S numbers. I could then determine whether it was is shade or not, by checking the V number. As long as the H and S numbers had remained constant, the pixel was likely to have changed colour due to being cast under shade, or being exposed to sunlight.

Left & Right: Sample attempts at shade detection indicated by the small green boxes.

2. Cumulative Change

Some areas of an image would show frequent amounts of change, whereas others remained constant. The cumulative change method involved remembering which pixels change a lot, and which pixels do not. If there was a large change in an area that changes frequently, then it is less likely that it was due to an animal. However if there is a small change in an area that never changes, then it is more likely to have been due to an animal. This method proved to be so effective, it often detected animals even when human observers had missed them!

Left: Original image. Right: Cumulative change mask image after applying animal detection on the original image. Animals clearly detected.

3. Box Histograms

This method involved dividing the image into small 20x20 pixel boxes. Each box had its own histogram which tallies up colours that it has seen. If a few rare colours appear in the box, then the new colours are tallied in the histogram and the box is declared as changed. If the colours are commonly seen (as shown by the histogram) they are simply added to the tally. This was the method I predominantly used as it was the most effective.

Left: Final user image. Right: Box histogram image after applying animal detection. The areas with the red boxes detected new colours and declared them as changed.

Conclusions

The methods I have used in this project have assisted in classifying images into those which have animals and those which do not. Each method performed well to address the different challenges I faced, however it was the combination of these three methods that yielded more effective and accurate results. The effect of shade continues to be a problem, indicating that more research in this area needs to be carried out.

[Top of Page]