基于光流法与特征统计的鱼群异常行为检测

鱼类群体行为的异常检测能够为鱼类健康监控与预警提供重要的方法和手段,对研究鱼类行为的机理,提升水产养殖过程中的信息化水平具有非常重要的意义。该文通过计算机视觉和图像处理技术,基于鱼群运动特征统计方法,对鱼群异常行为检测进行研究。利用Lucas-Kanade光流法得到目标鱼群的运动矢量,并对目标运动的行为特征进行统计,得到速度与转角这2个行为特征的联合直方图与联合概率分布。最后,在联合概率分布的基础上,基于标准互信息(normalized mutual information-NMI)和局部距离异常因子(local distance-based outlier factor-LDOF)2种方法对鱼群行为进行异常检测。试验结果表明,2种异常检测方法均达到99.5%以上的准确率。

Anomaly detection of fish school behavior based on features statistical and optical flow methods

Abstract: The behavior of fishes is very sensitive to the changes of the parameters of the environment, such as temperature, dissolved oxygen, light, and so on. The anomaly detection of fish school behavior can not only discover the relationship between the fish behaviors and the environmental parameters, but also provide an important method and tool for fish health monitoring and early warning. Moreover, it is very meaningful for the study of the mechanism of fish behavior and promotion of the informatization level in aquaculture. By using computer vision technology and based on a statistical method of motion features, the anomaly detection of fish school behavior was studied. The zebra fish was selected as the study object in this paper. First, based on the foreground object detection method with a threshold value method, the backgrounds were removed from the original video images to reduce the influence of noises. Secondly, by the Lucas-Kanade optical flow method, which is based on the local deference method and has better performance, the vectors of motion behavior could be obtained in different temporal and spatial conditions. Thirdly, from these data, the joint histograms and joint probability distributions of turning angles and velocities were calculated. Since from the practical point of view, the anomalous behaviors of a fish school mainly include the change of the moving velocity and the chaos of the moving direction. This is the reason to select turning angle and velocity as the features to analyze. At last, the NMI method and the LDOF methods were applied to study the anomaly detection of fish school behavior. By choosing proper threshold values, the NMI method and the LDOF methods can implement the behavior detection of the zebra fish school. The experiments showed that the accuracy rates of the NMI method and the LDOF method for anomaly detection of fish school behavior can achieve 99.92% and 99.88%, respectively, which implies that both of the two methods have better effects.