Drone- and AI-assisted livestock monitoring
| Reference number | |
| Coordinator | RISE Research Institutes of Sweden AB - ACREO, Kista |
| Funding from Vinnova | SEK 500 000 |
| Project duration | November 2018 - November 2019 |
| Status | Completed |
Important results from the project
The goal of the feasibility study was to verify which identification accuracy can be achieved using drone and AI-assistent monitoring of livestock. The project created and labelled over 700 RGB images of highland cows at a farm in mid-Sweden which have been used for training and testing with a version of Faster R-CNN algorithm adapted for the project purposes. The project achieved the average recall value of 89%, i.e. the algorithm found 9 of 10 cows on average. The algorithm also framed each identified animal on a picture with a box providing exact coordinates for positioning.
Expected long term effects
A bank of 700 labelled high-quality RGB drone images of highland cows. An extra bank of ca 500 RGB and 500 IR unprocessed images. Very good algorithm performance (90% recall for cows) even for images scaled down to 800x600 (see next section for clarification). The project came with an idea of a special maneuver or “fly by” at different heights that can help reducing the number of missed animals. The project has been invited to present its results at Data Innovation Summit 2020 in Stockholm in March 2020.
Approach and implementation
The project early identified candidate algorithms for performing the recognition of the animals at the terrain background. Faster R-CNN has been adapted for using in the project. The algorithm became very slow with images whose resolution was better than 800x600. The AI performance can thus be further improved when using full-resolution images. The algorithm has been further modified for use in the RISE SICS Hops platform for finer tuning of the hyperparameters and extracting of the performance metric such as recall and precision.