| 基于YOLO v8n-seg和改进Strongsort的多目标小鼠跟踪方法 |
| |
| 引用本文： | 梁秀英,贾学镇,何磊,王翔宇,刘岩,杨万能.基于YOLO v8n-seg和改进Strongsort的多目标小鼠跟踪方法[J].农业机械学报,2024,55(2):295-305，345. |
| |
| 作者姓名： | 梁秀英 贾学镇 何磊 王翔宇 刘岩 杨万能 |
| |
| 作者单位： | 华中农业大学 |
| |
| 基金项目： | 国家自然科学基金项目（U21A20205）、湖北省洪山实验室重大项目（2022hszd024）和山东省重点研发计划项目（2022CXGC010609） |
| |
| 摘 要： | 多目标小鼠跟踪是小鼠行为分析的基本任务，是研究社交行为的重要方法。针对传统小鼠跟踪方法存在只能跟踪单只小鼠以及对多目标小鼠跟踪需要对小鼠进行标记从而影响小鼠行为等问题，提出了一种基于实例分割网络YOLO v8n-seg和改进Strongsort相结合的多目标小鼠无标记跟踪方法。使用RGB摄像头采集多目标小鼠的日常行为视频，标注小鼠身体部位分割数据集，对数据集进行增强后训练YOLO v8n-seg实例分割网络，经过测试，模型精确率为97.7%，召回率为98.2%，mAP50为99.2%，单幅图像检测时间为3.5ms，实现了对小鼠身体部位准确且快速地分割，可以满足Strongsort多目标跟踪算法的检测要求。针对Strongsort算法在多目标小鼠跟踪中存在的跟踪错误问题，对Strongsort做了两点改进：对匹配流程进行改进，将未匹配上目标的轨迹和未匹配上轨迹的目标按欧氏距离进行再次匹配；对卡尔曼滤波进行改进，将卡尔曼滤波中表示小鼠位置和运动状态的小鼠身体轮廓外接矩形框替换为以小鼠身体轮廓质心为中心、对角线为小鼠体宽的正方形框。经测试，改进后Strongsort算法的ID跳变数为14，MOTA为97.698%，IDF1为85.435%，MOTP为75.858%，与原Strongsort相比，ID跳变数减少88%，MOTA提升3.266个百分点，IDF1提升27.778个百分点，与Deepsort、ByteTrack和Ocsort相比，在MOTA和IDF1上均有显著提升，且ID跳变数大幅降低，结果表明改进Strongsort算法可以提高多目标无标记小鼠跟踪的稳定性和准确性，为小鼠社交行为分析提供了一种新的技术途径。
|
| 关 键 词： | 小鼠行为 多目标跟踪 YOLO v8n-seg Strongsort |
| 收稿时间： | 2023/7/15 0:00:00 |
Multi-object Mice Tracking Based on YOLO v8n-seg and Improved Strongsort |
| |
| LIANG Xiuying,JIA Xuezhen,HE Lei,WANG Xiangyu,LIU Yan,ANG Wanneng.Multi-object Mice Tracking Based on YOLO v8n-seg and Improved Strongsort[J].Transactions of the Chinese Society of Agricultural Machinery,2024,55(2):295-305，345. |
| |
| Authors: | LIANG Xiuying JIA Xuezhen HE Lei WANG Xiangyu LIU Yan ANG Wanneng |
| |
| Institution: | Huazhong Agricultural University |
| |
| Abstract: | Multiple-object tracking of mice is a fundamental task in behavioral analysis and an important method for studying social behavior. In response to the limitations of traditional mouse tracking methods, such as the ability to track only a single mouse and the need for mouse labeling to track multiple-object mice, which affects mouse behavior, an unlabeled multiple-object mice tracking method was proposed based on the combination of instance segmentation network YOLO v8n-seg and improved Strongsort. RGB cameras were used to capture daily behavior videos of multiple-object mice, and a dataset for segmenting mouse body parts was annotated. After augmenting the dataset, the YOLO v8n-seg instance segmentation network was trained. The model achieved a precision of 97.7%, recall of 98.2%, mAP50 of 99.2%, and single-image detection time of 3.5ms. It accurately and quickly segmented mouse body parts, meeting the detection requirements of the Strongsort multi-object tracking algorithm. To address tracking errors in the Strongsort algorithm for multiple-object mice tracking, two improvements were made. Firstly, the matching process was improved by re-matching trajectories that did not match objects and unmatched objects based on Euclidean distance. Secondly, the Kalman filter was improved by replacing the rectangular bounding box representing the mouse position and motion state in the Kalman filter with a square box centered on the centroid of the mouse body contour and with a diagonal equal to the mouse body width. After testing, the improved Strongsort algorithm showed an ID switches of 14, MOTA of 97.698%, IDF1 of 85.435%, and MOTP of 75.858%. Compared with the original Strongsort, the ID switches count was reduced by 88%, MOTA was improved by 3.266 percentage points, and IDF1 was improved by 27.778 percentage points. Compared with Deepsort, ByteTrack, and Ocsort, there was a significant improvement in MOTA and IDF1, and the ID switches was greatly reduced. These results indicated that the improved Strongsort algorithm can enhance the stability and accuracy of unlabeled multiple-object mouse tracking, providing a technical approach for analyzing social behavior in mice. |
| |
| Keywords: | mouse behavior multi-object tracking YOLO v8n-seg Strongsort |
|
| 点击此处可从《农业机械学报》浏览原始摘要信息 |
| 点击此处可从《农业机械学报》下载免费的PDF全文 |
|
