基于改进YOLO v8s + DeepSort的奶山羊发情行为识别与跟踪方法

Estrus Behavior Recognition and Tracking Method of Dairy Goats Based on Improved YOLOv8s and DeepSort Algorithm

摘要:

奶山羊发情行为的准确识别和跟踪对提高养殖繁殖效率和管理水平至关重要。针对传统养殖对奶山羊发情期缺乏准确、及时的行为分析与监控方法问题，提出了一种基于改进YOLOv8s+DeepSort的奶山羊发情行为识别与跟踪方法。采用轻量化、无参数的SimAM注意力机制增强YOLOv8s的特征提取能力，并将其旋转边界框模型(YOLOv8s_obb)、常规矩形框模型(YOLOv8s)分别标记并检测奶山羊的爬跨行为和非爬跨行为；检测得到的行为类别信息和边界框参数信息与DeepSort跟踪算法结合，在其IoU匹配过程中引入匹配帧数计数机制，增强爬跨与被爬跨奶山羊之间的关联性，精准确定处于发情状态的奶山羊，获取对应的ID，确保发情行为在连续图像帧中能够准确跟踪。研究结果表明，改进模型对奶山羊爬跨行为和非爬跨行为的平均精度较YOLOv8s_obb和YOLOv8s分别提升了0.90、2.64个百分点；在发情山羊跟踪性能方面，与DeepSort、BotSort、StrongSort和ByteTrack相比，本文方法的HOTA和IDF1分数最高，分别达到72.4%、80.3%，可为奶山羊的繁殖管理提供有力支撑。

Abstract:

In modern dairy goat farming, accurate recognition and tracking of estrus behavior is crucial for improving breeding efficiency and management standards. Aiming at the lack of accurate and timely behavior analysis and monitoring methods for the estrus of dairy goats in traditional breeding, an estrus behavior recognition and tracking method of dairy goats was proposed. Firstly, the lightweight SimAM attention mechanism was used to enhance the feature extraction capability of YOLOv8s, and a rotating bounding box model (YOLOv8s_obb) and a conventional rectangular box model (YOLOv8s) were used to detect the mounting and the no mounting dairy goats, respectively. Then, combining the detected behavior category information and bounding box parameter information with the DeepSort tracking algorithm, the matching frame count mechanism was introduced into IoU matching process to enhance the correlation between the mounting and the no mounting dairy goats, and accurately determined the dairy goats in estrus, and obtained stable ID, which can ensure the accurate tracking of estrus behavior in successive image frames. The results showed that the AP of the improved model for detecting mounting and non-mounting behaviors in dairy goats was improved by 0.90 percentage points and 2.64 percentage points compared with YOLOv8s_obb and YOLOv8s, respectively. As for estrus goat tracking performance, compared with DeepSort, BotSort, StrongSort and ByteTrack, the proposed method achieved the highest HOTA and IDF1 scores, reaching 72.4% and 80.3% respectively, which can provide strong support for the reproductive management of dairy goats.