Feeding behavior is one of the most important influencing factors on large-scale pigeon farming for meat. More research has focused on target tracking algorithms in recent years. However, less attention has been paid to precise statistics of animal-specific behaviors using target tracking. In addition, some challenges are still remained on the behavior analysis in the small target scene of caged pigeons using machine vision, such as sparse target pixels, low resolution, overlap, obstruction by cage fences, complex background environment, weak lighting conditions, and image blurring. The target tracking algorithm is often required to accurately recognize the feeding behavior of pigeons, but resulting in missing or false detections and ultimately false trajectories. In this work, a precise statistical framework was proposed to detect the pigeon feeding frequency using a feeding behavior tracking algorithm and feeding trajectory, termed Mask-SCBF (SPD-Conv-BiFormer) -SORT. Three modules included: segmentation mask conversion, tracking algorithm, and feeding frequency calculation. A segmentation mask conversion module was introduced to convert into a bounding box and a score. The inputs were then selected for the tracking model in the tracking trajectory of the pigeon during feeding. The trajectory coordinates were extracted from the meat pigeon. A feeding frequency calculation was established using trajectory motion and a dynamic threshold. The results showed that: (1) A segmentation mask transformation framework was designed using an adaptive filtering mechanism. Feeding pigeon targets were extracted to improve the accuracy and robustness of segmentation; (2) A target tracking algorithm (SCBF-SORT) was proposed using the SCBF-YOLOv8 detector. In the tracking module, the model tracking accuracy was enhanced to reconstruct the Kalman filter state vector, and then optimize the noise covariance matrix parameters. In the detection module, an SCBF-YOLOv8 (SPD-Conv-BiFormer-YOLOv8) detector was designed with the SPD-YOLOv8 to enhance the feature representation of small targets using YOLOv8. In addition, the detector also included the following improvements. BiFormer attention mechanism was integrated to enhance contextual semantic correlation. WIoU v3 loss function was used to improve bounding box regression accuracy. A small object detection layer was added for the adaptability of target size. (3) Accurately counting feeding frequency was constructed using the target tracking trajectory and a dynamic threshold. Furthermore, over 2700 original videos were collected under three lighting conditions: strong light (daylight), weak light (nighttime illumination), and black light (infrared night vision). Key frames of feeding behavior were extracted from the original videos. A total of 4386 images with a resolution of 1920×1080 pixels were produced as the training and validation sets. Additionally, 15 videos with a relatively concentrated occurrence of feeding behavior, each 60 minutes long, were selected for the target tracking task. The experimental results showed that target behavior tracking was improved by 7.15%, 4.17%, 11.24%, and 5.81% in MOTA, IDF1, HOTA, and MOTP, respectively, with the target segmentation mask as the input, compared with the original images. The segmentation mask conversion can provide the target area information for the high accuracy of target tracking. The Mask-SCBF-SORT algorithm was improved by 7.52%, 5.36%, 14.39%, and 6.41% in the MOTA, IDF1, HOTA, and MOTP, respectively, compared with the baseline algorithm BoT-SORT. Small target detection and tracking performed best in complex environments of captive pigeons for meat. The high accuracy of feeding frequency reached 94.82%, fully meeting the demand for the rapid statistics of meat pigeon feeding behavior in actual breeding.