arrow drop search cross

A wolf counting experiment could lead to new methodologies in artificial intelligence

Publication : 29/07/2020
Partager cet article :

In the Mercantour Park, camera traps are hidden in tree hollows. The cameras are triggered by motion sensors in the hope of catching wolves passing by and ultimately of describing the populations present on the site more accurately. The number of images and videos collected for analysis is already considerable, which is why a partnership has been concluded with the 3IA Côte d'Azur institute and the Center for Modeling, Simulation and Interactions (MSI) to work together on processing the information. The first step will be to design artificial intelligence tools capable of detecting wolves in the database and of counting the species. At this stage, the problem is to deal with images acquired in a variety of conditions since the traps can be triggered at any time, in a wide variety of weather conditions, and the subjects can be captured at different angles and speeds. “The first thing we asked our MSI engineer was to make an assessment of the task involved considering how heterogeneous and rare the images are. Because there are not that many wolves after all! We expect this to be the most difficult problem,” explains Pr. Charles Bouveyron, deputy scientific director of the 3IA Côte d'Azur Institute, head of the Inria MAASAI team and recipient of a chair in Artificial Intelligence, which allows him to devote more time and resources to research in “statistical learning from heterogeneous data”.

For the “counting and detection” aspect of the project, which involves mostly applied research, the team is planning to use tools that have already become standard. The engineer appointed to the partnership will have to modify algorithm networks pre-trained to differentiate a wide range of objects in a very large database. This technique, called transfer learning, does not require the scientists to develop the artificial intelligence from scratch. “We will only need to touch the last layers of the network to adapt it very specifically to our situation,” summarizes Charles Bouveyron. “Now, we are waiting to find out on which level we will need to focus our efforts and use all the latest technologies or whether a network of basic deep learning will be sufficient. What worries us is not so much the detection part but rather the counting. We can sometimes find 26 chamois in a single sequence, passing quickly and sometimes very far away,” explains the researcher. But once the tools are operational, research will be much more theoretical. “The most ambitious aspect of the project that will require methodological and algorithmic developments will be to cross spatio-temporal data on the passage of the wolf with a wide range of information to achieve a very broad model of the activity of one species (the wolf) or of several species,” adds Charles Bouveyron. The camera traps allow researchers to know where and when the images were taken, but the guards of the Mercantour Park have also recorded animal footprints similar to GPS tracks on sections of the animals’ routes. They have also collected the results of a large number of genetic tests carried out on prey and droppings found in the Park. All of this information is combined to trace the animals individually.

“So far, there is no known methodology that could take all these characteristics into account. What is highly anticipated today from a theoretical perspective in machine learning or statistical learning is the capacity to process different kinds of data. In this case, we have the image, the GPS tracks and the genetic data,” Charles Bouveyron notes with satisfaction. “I like to be motivated in my work by the final application, but I know that the problem I am grappling with is much broader. The methodology that will be used here can be applied very widely.” The project covers two of the four strategic research themes of the 3IA Côte d'Azur Institute: fundamental artificial intelligence and the development of intelligent and secure territories. “We are working on something that is important for our region, that is applicable and is interdisciplinary. If we manage to complete the project in good conditions, which I hope will be the case, we will have achieved something very unusual. To my knowledge, we will be one of the only 3IA to carry out collaborative work with a National Park,” points out the deputy scientific director of 3IA Côte d'Azur, who is enthusiastic about “using Artificial Intelligence for the benefit of everyone”.