基于3D CNN-BiLSTM-ATFA网络和步态特征的奶牛个体识别方法

Individual Identification Method of Cows Based on 3D CNN-BiLSTM-ATFA Network and Gait Feature

摘要:

针对基于花纹的奶牛个体识别中纯色或花纹较少的奶牛识别准确率较低的问题，本文提出一种基于步态特征的奶牛个体识别方法。首先，将DeepLabv3+语义分割算法的主干网络替换为MobileNetv2网络，并引入基于通道和空间的CBAM注意力机制，利用改进后模型分割出奶牛的剪影图。然后，将三维卷积神经网络（3D CNN）和双向长短期记忆网络（BiLSTM）构建为3D CNN-BiLSTM网络，并进一步集成自适应时间特征聚合模块（ATFA）生成3D CNN-BiLSTM-ATFA奶牛个体识别模型。最后，在30头奶牛的共1242条视频数据集上进行了奶牛个体识别实验。结果表明，改进后DeepLabv3+算法的平均像素准确率、平均交并比、准确率分别为99.02%、97.18%和99.71%。采用r3d_18作为3D CNN-BiLSTM-ATFA的主干网络效果最优。基于步态的奶牛个体识别平均准确率、灵敏度和精确度分别为94.58%、93.47%和95.94%。奶牛躯干和腿部不同部位进行加权特征融合的个体识别实验表明识别准确率还可进一步提高。奶牛跛足对步态识别效果影响较为明显，实验期间由健康变为跛足和一直跛足的奶牛个体识别准确率分别为89.39%和92.61%。本文研究结果可为奶牛的智能化个体识别提供技术参考。

Abstract:

Aiming at the low identification accuracy of cows with solid color or less pattern in pattern-based individual identification of cows, an individual identification method was proposed based on cow gait features. Firstly, the backbone network of DeepLabv3+ semantic segmentation algorithm was replaced by MobileNetv2 network. The channel and space based CBAM attention mechanism was introduced into this segmentation algorithm. The improved model was used to segment the silhouette of the cow. Then the 3D convolutional neural network (3D CNN) and the bidirectional long short-term memory network (BiLSTM) were constructed as the 3D CNN-BiLSTM network. The adaptive temporal feature aggregation module (ATFA) was further integrated into the above network to generate the 3D CNN-BiLSTM-ATFA cow individual identification model. Finally, individual identification experiments were conducted on a total of 1242 video datasets from 30 cows. The results showed that the MPA, MIOU and Accuracy of the improved DeepLabv3+ algorithm were 99.02%, 97.18% and 99.71%, respectively. Individual recognition was optimal when r3d_18 was used as the backbone network of 3D CNN-BiLSTM-ATFA. The average accuracy, sensitivity and precision of individual identification based on cow gait were 94.58%, 93.47% and 95.94%, respectively. Individual identification experiments with weighted feature fusion for torso and legs showed that identification accuracy can be further improved. Lameness in dairy cows had a significant effect on gait identification, the individual identification accuracies were 89.39% and 92.61% for cows that changed from healthy to lame and cows that remained lame during the experiment, respectively. The results can provide technical reference for intelligent individual identification of dairy cows.