关键词: 笼养肉鸭, 多目标追踪, Mask R-CNN, 饮水采食, 行为识别

Abstract: Tracking and identifying the drinking and feeding behavior of caged meat ducks is of great significance to distinguish the pathological state of meat ducks and realize the intelligent management of duck houses. Aiming at the shortcomings of the traditional instance segmentation model that cannot correlate the logical sequence of video frames, a duck drinking and feeding behavior recognition model is proposed that first performs multiobject tracking （MOT） and then performs instance segmentation. A meat duck cage overhead shooting test platform was set up to collect image data containing multiple meat duck targets. The TAM （Track Anything Model） model was used to identify and track individual meat ducks. Then based on the SAM （Segment Anything Model）, the drinking and feeding behaviors of the three meat duck targets were quickly marked. Based on the multitarget tracking of meat ducks, Mask R-CNN was used to identify the drinking and feeding behaviors of ducks, and the duration of the two types of behaviors in the meat duck video was inferred based on the video frame rate, and finally the recognition and timing models of drinking and feeding behavior of meat duck was constructed. The test results show that the Mask R-CNN model has an average prediction frame accuracy rate  and mask average accuracy rate of the target meat duck's drinking water and feeding behavior recognition by 91.6% and 93.3%, respectively, the calculation accuracy of drinking water and feeding behavior duration is 95.4% and 90.1%, respectively, which can realize the recognition and timing of meat duck drinking and feeding behavior with high precision.

Key words: caged meat duck, multiple object tracking, Mask R-CNN, drinking and feeding, behavior recognition