49 / 2025-04-09 13:52:16

Behavior recognition and localization of cage-free chickens in videos based on spatiotemporal feature learning

Chicken; Behavior recognition; Behavior localization; Spatiotemporal feature; Computer vision.

3. 智慧牧场与动物福利 > （3）智慧养殖的传感器、智能感知和机器人

Timely access to chicken behavioral information can improve their welfare and reduce disease spread. Video-based behavior recognition methods have emerged as a primary technique for obtaining such information due to their accuracy and robustness. Video-based models generally predicted a single behavior from a single video segment of a fixed duration. However, during periods of high activity in poultry, behavior transition may occur within the video segment, and existing models often failed to effectively capture such transition. This study proposed an end-to-end method for recognizing multiple simultaneous chicken behavioral events and locating their behavioral boundaries in video segments by Chicken Behavior Recognition and Localization System (CBRLS) based on spatiotemporal feature learning. The CBRLS consisted of three main components: the improved YOLOv8s detector, the Bytetrack tracker, and the CBLFormer model. The basic network YOLOv8s was improved with MPDIoU to identify multiple chickens in the same frame of videos. The Bytetrack tracker was used to track each identified chicken and acquire its image sequence of 32 continuous frames as input for the CBLFormer model. To accurately recognize the behavior of each tracked chicken and locate its behavioral boundaries, the CBLFormer integrated an improved transformer block, a Cascade Encoder-Decoder Network (CEDNet), and a transformer-based head. For the training and testing of each component of CBRLS, the datasets were created by collecting videos from 320 chickens across different ages and rearing densities. The results demonstrated that the mAP@0.5 of the improved YOLOv8s detector was 99.50%. The Bytetrack tracker achieved a mean MOTA of 93.89% at different levels of occlusion. The CBLFormer achieved an mAP@0.5:0.95 of 98.34%. Furthermore, visualization results confirmed that the CBRLS correctly recognized chicken behavior categories and effectively captured the behavioral boundaries even when poultry behavior transitions occurred within the image sequence received from the tracker. This study provides an efficient tool for automatically and accurately recognizing chicken behaviors and locating their behavioral boundaries in videos.