Behavior analysis of broilers using computer vision technology is emerging as a core technology for managing health, welfare, and productivity in smart farms. However, in actual broiler house environments, existing object detection models face limitat...
Behavior analysis of broilers using computer vision technology is emerging as a core technology for managing health, welfare, and productivity in smart farms. However, in actual broiler house environments, existing object detection models face limitations in securing stable tracking data due to extreme occlusion caused by high rearing density and facilities, as well as appearance ambiguity among individuals.
To address these issues, this study proposes YOLO-DenseTrack, a novel fusion framework combining the YOLO (You Only Look Once) object detection model and a density map estimation model. The proposed system mutually complements YOLO's precise localization capability with the density map's ability to estimate the global count and identify dense regions.
Specifically, this study designed a bidirectional correction process using YOLO detection results and density map information. Areas identified as hotspots with high object probability in the density map but missed by YOLO are classified as "candidate areas," and re-detection with a lower threshold is performed to correct false negatives. Simultaneously, false positives from YOLO are removed based on the density values of the density map, thereby enhancing detection reliability.
Experimental results demonstrated that the proposed fusion model achieved significant improvements in Recall and Precision compared to the baseline YOLO model, proving superior detection performance especially in high-difficulty scenarios where objects are heavily overlapped or occluded. Furthermore, applying the enhanced detection results to a multi-object tracking algorithm (ByteTrack) improved tracking performance metrics such as MOTA and IDF1 Score. This study is expected to contribute to improving the accuracy of individual broiler behavior analysis and welfare monitoring systems by securing robustness in detection and tracking within high-density environments.