Feeding behavior is one of the critical welfare indicators of broilers. Hence, understanding feeding behavior can provide important information regarding the usage of poultry resources and insights into farm management. Monitoring poultry behaviors is typically performed based on visual human observation. Despite the successful applications of this method, its implementation in large poultry farms takes time and effort. Thus, there is a need for automated approaches to overcome these challenges. Consequently, this study aimed to evaluate the feeding time of individual broilers by a convolutional neural network-based model. To achieve the goal of this research, 1500 images collected from a poultry farm were labeled for training the You Only Look Once (YOLO) model to detect the broilers’ heads. A Euclidean distance-based tracking algorithm was developed to track the detected heads, as well. The developed algorithm estimated the broiler’s feeding time by recognizing whether its head is inside the feeder. Three 1-min labeled videos were applied to evaluate the proposed algorithm’s performance. The algorithm achieved an overall feeding time estimation accuracy of each broiler per visit to the feeding pan of 87.3%. In addition, the obtained results prove that the proposed algorithm can be used as a real-time tool in poultry farms.

This research proposes a low-cost system consisting of a hardware setup and a deep learning-based model to estimate broiler chickens’ feed intake, utilizing audio signals captured by piezoelectric sensors. The signals were recorded 24/7 for 19 consecutive days. A subset of the raw data was chosen, and events were labeled in two classes, feed-pecking and non-pecking (including singing, anomaly, and silence samples). Next, the labeled data were preprocessed through a noise removal algorithm and a band-pass filter. Then, the spectrogram and the signal envelope were extracted from each signal and fed as inputs to a VGG-16-based convolutional neural network (CNN) with two branches for 1D and 2D feature extraction followed by a binary classification head to classify feed-pecking and non-pecking events. The model achieved 92% accuracy in feed-pecking vs. non-pecking events classification with an f1-score of 91%. Finally, the entire raw dataset was processed utilizing the developed model, and the resulting feed intake estimation was compared with the ground truth data from scale measures. The estimated feed consumption showed an 8 ± 7% mean percent error on daily feed intake estimation with a 71% R2 score and 85% Pearson product moment correlation coefficient (PPMCC) on hourly intake estimation. The results demonstrate that the proposed system estimates broiler feed intake at each feeder and has the potential to be implemented in commercial farms.

Computer vision is one of the most scalable sensing modalities for welfare monitoring in commercialbroiler production, yet progress is constrained by a shortage of large, consistently annotated, openlyavailable datasets with shared evaluation protocols and baselines. ChickenVerse is presented here,an openly released dataset that aggregates imagery and video from five poultry farms spanningcontrolled research pens and full-scale commercial houses, several commercial broiler strains, obliqueand overhead camera viewpoints, and a range of resolutions, frame rates, and lighting regimes.ChickenVerse comprises two complementary datasets that share a common design and a welfare-oriented label vocabulary. ChickenDet provides 6,539 images with 153,764 annotated bird instancesin the Microsoft Common Objects in Context (MS COCO) format, each carrying a mutually consistentbounding box and pixel-level mask for object detection and instance segmentation. ChickenActprovides 15,250 short single-bird video clips spanning fifteen broiler behavior classes for videoclassification. Both datasets ship with recommended training, validation, and test splits and anevaluation protocol; the detection and segmentation splits are partitioned by source so that test scoresmeasure transfer to a previously unseen commercial house. Reference baselines are established forall three tasks: object detectors reach up to 83.7% average precision (AP) with RT-DETR-L across adocumented accuracy-versus-speed range; a Mask R-CNN attains about 77% mask AP, with fastersingle-stage YOLO-seg models trading accuracy for throughput; and a self-supervised V-JEPA 2model leads behavior classification with an 88.7% macro-averaged F1 score. The datasets, baselines,and accompanying code are publicly available at https://github.com/amirivojdan/ChickenVerse.