Deep Learning Based Animal Detection And Multiple-Object Tracking In Modern Livestock Farms

As the demand for consuming meat surges,livestock farmers need to use limited land to efficiently raise more animals.In the direction of digital farming,using images can extract various information about the animals in the farm to help farmers improve breeding efficiency.Among them,image-based animal detection and tracking have become a focus of researchers.In terms of animal detection,this article proposes to use the anchorfree detection network CenterNet for animal detection in farm images,and use the natural advantage of this network to obtain the position of animals to assist in animal behavior recognition.This architecture can directly extract the position and size of animals based on the semantic features of animal images.Compared with anchor-based architectures that require manual setting of anchor points,this architecture is more robust in farm scenes where animal sizes,postures,and quantities vary greatly.Through ablation experiments in farm scenes,this article proves that CenterNet can better balance the relationship between detection speed and detection accuracy.The detection results are more consistent with the actual position and size of the animals than the network based on anchor points,and the obtained animal positions can effectively use relevant prior information in the background of farm monitoring videos to improve the accuracy of animal behavior recognition.In the tracking task,this article considers the similarity of the position and appearance of the same animal between consecutive frames,and the differences from other animals,and proposes a multi-object tracking scheme that uses animal detection results and animal position and appearance semantic features as similarity criteria.Based on this,a center point prediction module is designed to use the continuity of animal position in the video to assist in detection and tracking.This scheme reduces the difficulty of modeling animal trajectories by performing animal detection on consecutive frames,and is more robust than traditional multi-object tracking schemes that directly model animal trajectories in real farm scenes.At the same time,experimental results on multiple farm scene datasets demonstrate that the center point prediction module can effectively improve the detection and tracking accuracy of many similar animals appearing in the same scene.