Huu Quy Nguyen, a student at the Faculty of Information Technology (FIT) of the Czech Technical University (CTU), focused on creating a recommendation system for optimizing the monitoring of live streams of elephants at the Prague Zoo in his bachelor’s thesis. This system helps viewers select the most suitable camera to watch the elephants and also eases the work of the zookeepers by automating the tracking of the elephants' locations throughout the day, a task that was previously done manually. This project was developed in collaboration with Prague Zoo, which had previously approached FIT for potential cooperation. As a result of this work, Huu Quy Nguyen was awarded the Dean's Award.
For the project, over 10,000 images were collected from eight cameras placed in various parts of the Elephant Valley complex. More than 13,000 elephant objects (images/silhouettes) were labeled on these images. Based on this data, an elephant detection model was trained using the YOLOv8 technology, which is highly accurate and can recognize elephants even under poor lighting conditions. A heatmap was created from the collected data, showing where the elephants are most frequently located. Using this information, a system was developed that recommends the optimal camera for watching the elephants. The model can correctly recommend a camera with the most elephants in 62% of cases and a camera that captures at least one elephant in 93% of cases, making it easier for viewers to select a camera where they are most likely to see at least one elephant.
“When searching for a topic, this project caught my attention from the first read because I was interested in computer vision, but most importantly, I wanted to work on a project that would have a direct positive impact,” says Huu Quy Nguyen. “The zookeepers at Prague Zoo were very pleased with the heatmaps showing elephant occurrences throughout the day, as they seemed to match their own experience,” he adds.
There was also discussion about automating the system for recording the elephants' positions throughout the day, which would greatly ease the work of the zookeepers who currently do this manually. However, the resulting recommendation system is not yet ready for deployment, as it is based on historical data spanning four months, and the rest of the year has not been recorded.
“It would be possible to add additional data and improve the recommendation system, which would then be very easy to implement,” adds Huu Quy Nguyen, who is currently on a study trip in Japan but would like to discuss the potential deployment with the zoo's management upon his return.
