基于YOLOX-BT的奶牛多目标跟踪方法

Multi-object Tracking Method for Cows Based on YOLOX-BT

摘要:

奶牛跟踪是获取奶牛位置信息和活动数据的重要方式,对奶牛的智能管理具有重要意义。针对视频中部分奶牛目标较小、存在重叠与环境遮挡造成的漏检、误检和轨迹断裂的问题,提出了YOLOX-BT 算法,该算法将 YOLOX-s的CSP2_1 结构替换成了CFP_EVCBlock结构,用于提升小目标检测性能;在CFP_EVCBlock结构中,通过整合多种特征增强机制设计了特征重标定模块(Feature refinement module,FRM),解决重叠与遮挡问题,实现了对重要特征的有效提取;为了改善轨迹断裂情况,在数据关联部分,引入外观相似度改进相似度计算方法,并且在匹配过程中,动态调整外观相似度匹配和IoU匹配的所占权重,提升了奶牛的匹配精度。实验结果表明,改进后的检测器平均检测精度、精确率、召回率、帧速率分别达到了94. 9% 、94. 2% 、92. 7% 和22. 9 f/ s,较原模型相比分别提高了1. 2、0. 8、2. 3 个百分点和29. 4% ,参数量下降了9. 3% ,浮点运算量减少了16. 7% 。YOLOX-BT方法在奶牛多目标跟踪任务上,MOTA为93. 2% ,IDF1 为95. 2% ,IDs 为13,MOTP 为5. 3% ,与目前常用方法相比,在MOTA和 IDF1上均有显著提升,并有效减少了ID切换。YOLOX-BT算法能实现养殖场奶牛的准确且持续有效跟踪,能够为奶牛智能管理提供技术支持。

Abstract:

Cow tracking is an essential technique for obtaining positional and behavioral data of cows, playing a vital role in intelligent farm management. To tackle challenges such as missed and false detections, as well as track fragmentation caused by small object sizes, occlusion, and overlap in farm videos, the YOLOX-BT algorithm was proposed. This method enhanced both detection and tracking performance. Specifically, the YOLOX-s model was improved by replacing the original CSP2 _ 1 structure with the CFP_EVCBlock, boosting small object detection capability. Within CFP_EVCBlock, a feature refinement module ( FRM ) was designed by integrating multiple feature enhancement mechanisms, enabling prioritized extraction of important features and effectively mitigating occlusion and overlap issues. In the tracking phase, appearance similarity was introduced into the data association process, and the weights between appearance and IoU matching were dynamically adjusted, improving cow identity matching accuracy. Experimental results showed that the improved detector achieves an average precision of 94. 9% , precision of 94. 2% , recall of 92. 7% , and an inference speed of 22. 9 f/ s, representing gains of 1. 2, 0. 8, 2. 3 percentage points, and 29. 4% , respectively, over the baseline. Additionally, the number of parameters and FLOPs were reduced by 9. 3% and 16. 7% , respectively. In the cow multi-object tracking task, YOLOX-BT achieved a MOTA of 93. 2% , IDF1 of 95. 2% , 13 ID switches, and a MOTP of 5. 3% . Compared with commonly used methods, the proposed algorithm showed significant improvements in both tracking accuracy and identity consistency, providing robust technical support for intelligent cow management.