基于自适应无参核密度估计算法的运动奶牛目标检测

Detection of Moving Cows Based on Adaptive Kernel Density Estimation Algorithm

摘要:

复杂养殖环境下运动奶牛目标的准确检测是奶牛跛行、发情等运动行为感知的基础。针对现有方法多采用参数化模型实现运动奶牛目标检测的缺陷，提出了一种无参核密度估计背景建模方法。该方法根据各像素的历史样本估计像素的概率模型，针对历史样本信息中冗杂信息导致模型复杂度过高的问题，采用关键帧检测技术剔除样本中的冗余信息以降低算法的复杂度，并实现了在小样本下核函数对遥远历史帧图像信息的获取，从而提高了检测精度。针对检测目标轮廓缺失的问题，结合三帧差法进一步实现了运动目标的完整提取。为了验证本算法的有效性，对不同环境和干扰下的运动奶牛视频样本进行了试验，并与高斯混合模型(Gaussian mixture model, GMM)和核密度估计模型(Kernel density estimation, KDE)方法进行了对比。试验结果表明，本文算法平均前景正检率为95.65%，比高斯混合模型提高了15.56个百分点，比核密度估计模型提高了10.56个百分点。同时，本文算法平均实时性指标为1.11，基本可以实现运动奶牛目标的实时、准确检测，该研究结果可为奶牛跛行疾病的预防、诊断以及奶牛运动行为的精确感知提供参考。

Abstract:

Realizing the accurate detection of moving cows in complex farming environment is the basis for behavioral perception of cows such as lameness detection and estrus behavior analysis. Aiming to solve the defects of the existing methods using parametric model to achieve the target detection of moving cows, a background modeling method based on nonparametric kernel density estimation was proposed. The probability model of the pixel was estimated according to the historical sample of each pixel, which had the disadvantage of high complexity caused by the redundant information contained in the historical sample information. The key frame detection technique was adopted to eliminate the redundant information in the sample to reduce the complexity of the algorithm and the ability to acquire the remote frame image information by the kernel function under small samples to improve the detection accuracy. In view of the lack of detected target contours, the threeframe difference method was applied to further achieve a more complete extraction of moving targets. In order to verify the effectiveness of the proposed method, the video samples of moving cows under different environments and disturbances were tested and compared with the Gaussian mixture model and the Kernel density estimation model. The experimental results showed that the average detection rate of the proposed algorithm was 95.65%, which was 15.56 percentage points higher than that of the Gaussian mixture model and 10.56 percentage points higher than that of the Kernel density estimation model. It also showed that the research algorithm had greater improvement than the Gaussian mixture model and the Kernel density estimation model in complex environments such as sunny, rainy and night time. In addition, the average realtime indicator of the algorithm was 1.11, which can basically realize the realtime and accurate detection of moving cow targets. The results were of great significance for the prevention and diagnosis of dairy cows disease and the accurate perception of cows movement behaviors.