This study aims to improve the efficiency of livestock production by automating the process of monitoring herds of farm animals under free-range grazing conditions. To this end, a methodology for the automated synthesis of a high-performance animal recognition model in a natural habitat has been developed using the AutoGenNet AutoML framework. During the study, relevant technologies were analyzed, the AutoGenNet framework was modified for video stream processing, and its integration with the external tracking modules DeepSORT and BOT-SORT was implemented based on experimental results. The effectiveness of the proposed methodology was evaluated using video data acquired by a DJI Mavic Pro unmanned aerial vehicle (UAV). Video recording of a cattle herd was performed on an open pasture under challenging conditions and with frequent occlusions (in computer vision terminology, cases in which some objects block others from the camera’s field of view). In the 104 s video sequence, 20 cows were recorded; the video acquisition conditions included dynamic camera motion (40 instances of animals leaving the frame followed by re-entry) and frequent occlusions (24 cases). Based on the experimental results, the best automatically synthesized system configuration produced 72 errors over the entire video duration. These errors were largely caused by the specifics of the video acquisition process and can be mitigated by modifying the data acquisition methodology in order to minimize occlusions and repeated re-entry of already counted animals into the frame. The proposed methodology is suitable for rapid development and validation of monitoring systems, as well as for selecting the most appropriate tracking component based on quantitative data. The approach simplifies the process of creating and integrating video monitoring system components, which reduces dependence on their manual tuning by machine learning specialists.
Sobolevsky et al. (Mon,) studied this question.