Solving the problem of biodiversity analysis of bird detection and classification in the video stream of camera traps

Abstract

The work is devoted to solving the problem of assessing the comparative efficiency of several common architectures of convolutional neural networks for monitoring birds in a natural environment. The problem was solved by detecting birds recorded by video traps installed on feeders in several regions of Panama by different architectures. Then a comparison was made between the recognition quality metrics – IoU and mAP, and based on the values of the metrics, a conclusion was made about the effectiveness of the architectures. Experiments have shown that the YOLO architecture of the Tiny version with comparative modules wins in the accuracy table. In the future, it is planned to improve the application of neural network architectures by finalizing the dataset with the involvement of expert bird watchers and open ornithological ontologies.