基于GMM-CNN的鱼道内鱼类游泳行为提取方法

鱼道设计离不开鱼类行为特性的研究，智能监控是观测、记录和量化鱼类行为的主要手段，但由于鱼道紊动的水流环境对鱼体精确跟踪造成了困难。本研究提出一种适用于鱼道研究的鱼类游泳行为提取方法，通过高斯混合模型（GMM）和卷积神经网络模型（CNN）对鱼体图像语义分割，再根据鱼体像素计算上溯轨迹、游泳速度和摆尾幅度。方法的验证以草鱼为试验对象，分别进行基于UNet和SegNet的鱼体语义分割对比试验和鱼类游泳行为提取试验。试验结果表明：（1） UNet模型的图像语义分割结果（MPA=95.9%，MIoU=93.3%）优于SegNet模型（MPA=95.5%，MIoU=92.8%）；（2）应用GMM-CNN模型进行过鱼轨迹跟踪，其坐标平均相对误差（MRE_x=7.1%，MRE_y=2.4%）低于单独使用GMM方法23.1%、11.6%；（3） GMM-CNN模型测算鱼类行为能够准确定位鱼体特征点，应用于提取鱼类游泳行为时具有较高的精确度。基于GMM-CNN鱼类游泳行为提取方法可以为鱼道设计提供技术支持。

Extraction method of fish swimming behavior in fishway based on GMM-CNN

Fishway design is inseparable from the study of fish behavior characteristics.Intelligent monitoring is the main means of observing, recording and quantifying fish behavior.However, accurate tracking of fish body is difficult due to the turbulent flow environment in fishway.This study proposes a fish swimming behavior extraction method suitable for fishway research, which uses Gaussian mixture model (GMM) and convolutional neural network model (CNN) to semantically segment the fish body image, and then calculates the upstream trace, swimming speed and tail wagging amplitude according to the fish body pixels.The comparison experiment of fish body semantic segmentation based on UNet and SegNet and the experiment of fish swimming behavior extraction were carried out with grass carp as the test object.The experimental results show that:(1) the image semantic segmentation results of UNet model (MPA=95.9%, MIoU=93.3%) are better than those of SegNet model (MPA=95.5%, MIoU=92.8%);(2) The average relative error of coordinates (MRE_x=7.1%, MRE_y=2.4%) was lower than that of GMM alone (23.1%, 11.6%);(3) The GMM-CNN model can accurately locate the characteristic points of the fish body and has a high accuracy when measuring the fish behavior.The extraction method of fish swimming behavior based on GMM-CNN can provide technical support for fishway design.