基于改进YOLO 11n-Pose和Jetson Orin NX的多目标小鼠行为分析方法

Multi-object Mice Behavior Analysis Method Based on Improved YOLO 11n-Pose and Jetson Orin NX

摘要:

针对人工观察小鼠行为存在费时费力、主观性强的问题，提出了一种适用于Jetson Orin NX平台的多目标小鼠行为分析方法。经比较，YOLO 11n-Pose在多个姿态关键点模型中效果最优，精确率为94.1%、召回率为94.5%、平均精度均值为96.4%，单幅图像推理用时36.04 ms。提出WIoU损失函数改进，精确率提升2.1个百分点，平均精度均值提升0.7个百分点。基于结构简化和量化模型并部署到Jetson Orin NX，压缩后模型的精确率、召回率和平均精度均值分别为93.93%、94.18%和95.88%，单幅图像推理用时11.01 ms，缩短69.46%。对比4个多目标跟踪模型，结果表明OC Sort的效果优于其他模型，单幅图像推理用时2.18 ms。基于OC Sort多目标跟踪模型进行改进，改进后的IDF1、MOTA和MOTP分别为83.5%、92.0%和27.6%，ID跳变比原模型减少41.18%, 单幅图像推理用时2.32 ms，结果表明改进OC Sort在不影响实时性的同时对于准确性有明显提升；采用基于阈值法的小鼠行为预测，将人工值和预测值对比，结果表明蜷缩、弯曲和伸长行为判断准确率分别达到95.12%、81.71%和80.49%；采用该方法对比青年鼠与老年鼠24 h的行为活动，结果表明青年鼠的活跃度明显高于老年鼠，青年鼠更多时候处于休息状态，而老年鼠则更多处于睡觉状态。研究表明，改进YOLO 11n-Pose和OC Sort相结合的多目标小鼠行为分析方法，在提升跟踪与关键点准确性的同时，经量化部署后可实现便携、实时的多目标小鼠行为分析。

Abstract:

Aiming to address the drawbacks of manual observation of mouse behavior, which is time-consuming, labor-intensive, and highly subjective, a multi-object mice behavior analysis method adapted for the Jetson Orin NX embedded platform was proposed. Through comparative experiments among multiple pose keypoint estimation models, YOLO 11n-Pose achieved the optimal comprehensive performance, with a precision of 94.1%, recall of 94.5%, the mean average precision was 96.4%, and inference latency of 36.04 ms per single image. To further optimize the model, an improvement scheme with WIoU loss function was proposed, which increased the precision by 2.1 percentage points and the mAP by 0.7 percentage points. Subsequently, model structure simplification and quantization were performed for edge deployment on the Jetson Orin NX platform. The compressed model achieved a precision of 93.93%, recall of 94.18%, and mAP of 95.88%, with the inference latency per single image reduced to 11.01 ms, a 69.46% reduction compared with that of the original model. Furthermore, comparative experiments were conducted on four mainstream multi-object tracking models, and the results showed that OC Sort outperformed other models with an inference latency of 2.18 ms per single image. On this basis, an improved OC Sort model was proposed. The optimized model achieved an IDF1 score of 83.5%, MOTA of 92.0%, and MOTP of 27.6%. The number of ID switches was reduced by 41.18% compared with that of the original OC Sort, while the inference latency was only 2.32 ms per single image. These results indicated that the improved OC Sort model significantly enhanced tracking accuracy without compromising real-time performance. For mouse behavior recognition, a threshold-based method was adopted. By comparing manual annotation results with model recognition outputs, the recognition accuracy of curving, shrinking and elongation behaviors reached 95.12%, 81.71% and 80.49%, respectively. The proposed method was applied to compare the 24 h behavioral activities of young mice and aged mice. The results showed that the activity level of young mice was significantly higher than that of aged mice; young mice stayed in a resting state for more time, while aged mice were mostly in a sleeping state. It demonstrated that the proposed multi-object mice behavior analysis method, which combined the improved YOLO 11n-Pose and optimized OC Sort models, can not only improve the accuracy of keypoint estimation and multi-object tracking, but also can realize portable and real-time multi-object mice behavior analysis after quantization and edge deployment.