Thermal Cameras vs Conventional Cameras for Predator Detection in NZ

Conventional trail cameras are cheap and offer high image resolutions. They are designed for detecting larger animals such as pigs and deer. Thermal cameras, like those used by the Cacophony Project, are much more expensive and typically have lower resolution. Conventional cameras need an IR light source to "see" at night while thermal cameras do not.

One of the reasons that trail cameras are so poor at tracking small animals is due to the basic sensors they use to detect motion. These work well for larger animals but often miss the smaller, fast-moving predators that are problematic in New Zealand.

The Cacophony Project has looked into whether it is possible to improve the detection of conventional cameras by running them in an "always-on" mode instead of relying on a motion sensor. After much testing why have opted to use always-on thermal cameras instead and here's why:

Thermal cameras are more sensitive at detecting NZ predators, despite their lower resolution.
It is much easier to apply machine learning classification to thermal camera footage.
Always-on conventional cameras use more power than thermal cameras because they need a constant IR light source at night.
It is likely that IR light is visible to and repels some kinds of introduced predators.
The relative cost of thermal cameras is high now but we think a thermal camera solution may end up being lower cost overall when compared to an equivalent conventional camera solution.
Thermal cameras have advantages at night, when most predators are active.

Let's go through these points in more detail.

Thermal cameras are more sensitive than conventional cameras

In 2016 The Cacophony Project started testing to compare the performance of thermal cameras and conventional cameras for predator detection. The first thermal camera we used was the 60 x 80 pixel FLIR Lepton 2. This was compared to a conventional 1280 x 720 resolution camera. Both the cameras were plugged into a Raspberry Pi and recordings compared. Full details of this test are available in an earlier blog post.

The following example video clearly shows how even a 60 x 80 thermal camera is better at detecting animals than a much higher resolution conventional camera.

We then performed extensive testing with a system that had both a thermal camera and conventional camera mounted side by side and again it is clear how much easier small predators are to detect using a thermal camera.

In 2017 we tried a FLIR Lepton 3 thermal camera module that has four times the resolution of the previous thermal camera model. The results were so compelling that we no longer saw the need to use the conventional camera alongside a thermal camera.

Later in 2017 we ran a number of preliminary tests that showed how much more sensitive a thermal camera is compared to a trail camera. These results have been confirmed by more recent experiments.

The summary of all this testing is that low resolution thermal cameras detect small animals with much greater sensitivity than conventional cameras.

Integrating Artificial Intelligence (AI) is easier with thermal cameras

The two parts of AI for predator detection are:

finding animal in a picture or video
classifying the animal

Locating animals in conventional video footage can be very challenging but with a thermal camera finding animals is almost easy as animals stand out against the relatively cool background. The videos here show raw thermal footage next to processed footage where the animal has been isolated. AI classification details are also shown. The numbers on the box at the top is the best guess and the number at the bottom is the real time calculation as the animal moves.

We already have our AI predator classifier working well enough now that it significantly reduces the human labour required for predator detection, resulting in a lower cost to operate than cheap chew cards.

AI based detection using conventional trail camera footage is fairly common practice but the success rate is typically higher for larger animals. There has been some success using AI with conventional cameras for NZ predators but it is worth noting that these cameras miss up to 95% of predators that come into view. This means that the AI is actually only working with the 5% easiest to detect predators!

Conventional always-on cameras use more power

There are two reasons for significant power consumption in a sensitive predator detection device:

The computer running the camera and detection needs to be on all the time. It should be noted that we believe we can significantly reduce the power used by the computer in the Cacophony Project's thermal camera (currently 1.5-2W) but this hasn't been a focus for us so far.. One electronics engineer has estimated that we could reduce our current power consumption by a factor of 10 by changing the hardware architecture used.
A conventional always-on camera operating at night needs a constant IR light source This is 1.5-2W for a lower light option and 3-4W for the higher brightness option. A thermal camera does not need any light source.

This means that the power consumption for an always-on conventional camera will be 2-3 times that of a thermal camera. We discuss in another article how it makes sense to use more power to improve detection rates. It isn't sensible to use more power on a tool that is less effective.

IR light may repel some predators

There is some evidence that the IR light can act as a deterrent to some parts of the predator population. Our goal is to monitor and eliminate 100% of introduced predators making conventional cameras a less desirable option for detection.

Cost of thermal camera solution likely to be lower long term

While the cost of thermal cameras is high now we think they may end up lower cost than equivalent conventional camera solution that uses more power.

Over time it is likely that power used by the compute component will reduce more quickly than the power used by the light source. This is likely to be even more pronounced for the thermal camera as it produces less output data which is less computationally intensive - and therefore power intensive - to handle.

The battery pack is a significant part of the cost of the product and historically has been on a more gradual cost reduction path than thermal camera chips. It seems likely that whatever is saved on a lower cost camera will be outweighed by larger battery costs long term.

Thermal cameras are better at night

Conventional cameras have some advantages during the day, particularly if used in sunny areas, but thermal cameras have many advantages at night (more sensitive, lower energy). Most of the predators we are interested in are primarily active at night so it makes sense to use a camera technology that works best at night.

Summary

Conventional cameras are cheap and high resolution. Thermal cameras are expensive and have lower resolution. Even so, extensive testing on New Zealand predators indicates that thermal cameras are more effective and likely to have a lower overall cost for tackling New Zealand's predator management challenges.