The National Tiger Conservation Authority in collaboration with the State Forest Departments, the Wildlife Institute of India (WII) and conservation partners conducts a national assessment for the “Status of Tigers, Co-predators, Prey and their Habitat” once in every four years. The aim of this exercise is to gauge the success of conservation efforts and to keep track of tiger populations and their ecosystems. What had until recently been a manual process of counting the individual tigers in the forests has now sought to be automated with the use of artificial intelligence techniques. 

It was in 2016, when Ankita Shukla was in her second year of PhD at the Indraprastha Institute of Information Technology, Delhi, under the supervision of Dr. Saket Anand, when she was offered to work on the tiger project with WII. The idea was to leverage computer vision for solving a real world problem. “I agreed to work on this project because I have a lot of affection for animals,” says Ankita, in her conversation with INDIAai.

It was a challenging research journey to embark upon. “At that point of time, we had no clue what kind of a data set it is going to be. But nevertheless, we started getting a lot of tiger images and these were all camera trap images,” says Ankita. But there was an inherent risk in sharing this data, which if acquired by bad actors, could be catastrophic for animal safety. “WII ensured that the images do not have location information or the path location.” 

“As we started looking at those images, we realised it's more challenging to work with tiger images as opposed to our standard computer vision datasets,” explains Ankita. One of the specific issues, when dealing with tiger images, was the camouflage problem. “It's very easy for a tiger to get blended in the background; and while we can see it, it's very difficult for a machine to differentiate. In fact, this can be a tedious task for humans too, sometimes. For example, if it's in a very dry landscape, then the rocks start looking somewhat like the tiger itself, at least for the machine if not for us.”

Another problem was making the distinction between different individuals to establish their uniqueness. “It’s very tough to tell apart one tiger from another because the difference is very subtle, and it exists only in their stripe patterns. So the question was, how can we make an AI system, which can use just a small amount of data to capture the fine-grained information. Since this requires very sensitive data, we ran our algorithms on publicly-available datasets and wrote a paper around it.” Their paper can be accessed here.

Ankita talks about another use of AI in the wild– a broader problem than even the tiger identification task – the detection of different animal species. “You can think of species classification as a pre-processing step to even the tiger identification problem because you’ll need to separate out the tiger images from rest of the species first, right?”

Giving an insight into the workings of the AI system, she explains, “So when you're talking about the motion triggered cameras, they're not going to differentiate between a tiger or bull or pig – they will capture everything. Sometimes, you even get human images, for example, if there's someone going on a motorbike for putting something in the park. The first task, therefore, was to segregate the tiger images from the rest because we need to work on individual identification. And the second task was to understand what the species categorisation is in a particular area.” The prelude to classification of species is, of course, the identification of all species in an area. “The task of this AI system that we devised was to identify the species present in a given a set of images",” says Ankita.

Co-developed with another researcher, Gullal Singh Cheema, this image processing software known as CaTRAT (Camera Trap Data Repository and Analysis Tool) is currently being used by WII for geotagging, coding and segregating the images to individual species folders. The geo-tagged images are scrutinized for potential software misclassification.

Image by JOSE ALMEIDA from Pixabay