基于三维姿态拟合的非接触式红鳍东方鲀全长精准估算方法

鱼类生物量无损测算是智能化水产养殖的重要环节，如何实现鱼体全长精准估算是该环节稳定运行的重要前提。该研究以红鳍东方鲀为对象，提出了一种鱼体全长精准估算方法，可在非接触情况下对自由游动的红鳍东方鲀进行精准的体长估算。首先，利用双目立体视觉技术对原始图像进行校正和立体匹配获得深度图像，并通过SOLOv2模型进行鱼体分割；然后，通过自主设计的独立分类器对图像进行高效分类，自动获取可用于全长估算的鱼类侧面图像，其分类准确率达95.3%；最后，耦合图像平面特征和深度信息，对鱼类进行三维姿态拟合，实现鱼类全长精准估算。结果表明，该方法全长估算的平均相对误差为2.67%，标准差为9.45%，且全长估算值与质量表现出良好相关性（R2=0.88）。该研究将为鱼类生物量无损测算提供关键技术支撑，对水产养殖的信息化管理、鱼类生长状况评估、投饵控制等具有重要意义。

Non-contact method for the accurate estimation of the full-length of Takifugu rubripes based on 3D pose fitting

Abstract: Non-destructive measurement of fish biomass is the key to achieving intelligent farming in aquaculture. Among them, the prerequisite can be the accurate estimation of the full length of fish. In this study, an accurate estimation was proposed for the full length of fish using 3D pose fitting. A summary was given on the previous estimations of fish length in recent years. Taking the Takifugu rubripes as the research object, the accurate estimation of the body length was realized in a non-contact situation. Firstly, the original image was corrected and stereo matched to obtain the depth image using binocular stereoscopic vision technology, and then the foreground segmentation of the fish body was performed by SOLOV2. Secondly, an independent classifier was designed to automatically acquire the fish side images for the full-length estimation. A comparison was made on the classification effects of three classifiers (SVM, Resnet18 and ResNet34). Finally, the ResNet34 network was selected as the image classification with the highest priority of accuracy, while maintaining a short inference time for the single image. As such, the accuracy reached 95.3%, and the inference time was 0.040s in the NVIDIA Geforce RTX 3090 SLI. Thus, the images were efficiently classified to improve the data quality, where the 2135 images of 121 fish were automatically filtered for the subsequent full-length estimation. The image data was fully used to couple the plane features and depth information of the fish images. 3D pose fitting of fish in the x-y and x-z planes was also performed to achieve the full-length estimation of fish. The results showed that the linearity fit was good with R2 of 0.90, indicating better linearity between the manually measured and the estimated total length. When the estimated total length and body weight were fitted with the multiplicative power formula, the fit was good with an R2 of 0.88, indicating a better correlation. The error analysis showed that the mean absolute error of the estimated full length was about -10.5mm, the mean relative error was 2.67%, the standard deviation of accuracy was 9.45%, and the relative error of the single full-length prediction data fell in the range of -6.68% to 12.12%, demonstrating a higher accuracy and stability. The estimated full length was also taken with the manually measured length as the exact value. Only image data was as the input, where the biological characteristics (such as body size and pigmentation) were the primary judgment factors for the applicability. The length was estimated during fish entering the developmental stage, where the biological characteristics were similar to those of adult fish. Meanwhile, the applicability range of this model for Takifugu rubripes was at least 187-650 mm full-length, according to the growth process and the applicability range of the multiplicative power formula in the existing studies. This finding can provide important technical support for the non-contact fish biomass estimation in the aquaculture process, particularly for the information management, fish growth condition assessment and baiting control of aquaculture.