基于ALR-GMM的群养猪攻击行为识别算法研究

Recognition of Aggressive Behaviour in Group-housed Pigs Based on ALR-GMM

摘要:

群养猪攻击行为是评估猪群对微环境适应性的重要指标。活动指数模型能够描述猪群行为模式，已经在群养猪攻击行为识别研究中得到初步验证。然而，养殖设施的差异性和动态背景环境等因素所导致的环境适应性差是限制其商业化应用的主要障碍。本文基于递归背景建模思想，在高斯混合模型（GMM）中引入双曲正切函数，提出了一种自适应学习率GMM的活动指数计算方法（ALR-GMM），能够在动态背景环境下准确提取动物活动指数。与经典模型相比，平均相对误差从15.08%降到14.34%。育肥猪攻击行为识别试验中，采用ALR-GMM算法提取行为视频单元的活动指数特征，构建了活动指数最大值、平均值、方差和标准差特征向量，采用线性核函数支持向量机建立分类器。结果表明，本文算法的正确率、灵敏度、特效度和精度分别为97.6%、97.9%、97.7%和97.8%，满足实际应用需求。

Abstract:

The activity index has been widely used as an important indicator to quantify the behavior response of livestock animals to their micro-environment. The variety of breeding facilities and dynamic backgrounds brings challenges to precisely monitor this indicator by the approach based on background subtraction. Aiming to extract the activity index under the dynamic background, a novel approach based on the Gaussian mixture model was proposed, in which the hyperbolic tangent function was introduced to regulate the learning rate parameter (ALR-GMM). In each iteration, through an adaptive learning rate regulation mechanism, the ALR-GMM can describe the image background and detect the moving pixels synchronously. The algorithm was evaluated on the manually labeled image. Compared with the calculation methods based on the background subtraction and the classical GMM methods, the mean relative errors were reduced by 0.74 percentage points and 3.74 percentage points, respectively. In order to further verify the feasibility, the proposed algorithm was applied to recognize aggressive behavior of group-housed pigs. The original video was divided into 3s episode units. In each video unit, the maximum, mean, variance and standard deviation of the activity index were taken as the feature vector. The behavior classifier was established by the linear kernel SVM. The results showed that the accuracy, sensitivity, specificity and precision were 97.6%, 97.9%, 97.7% and 97.8%, respectively, which met the requirements of practical application.