Development Of Non-Contact Cow Condition Scoring System And Equipment Based On Deep Learning

Body condition scoring for dairy cows is a method of assessing their obesity level.The body condition of cows that are too thin or too fat can significantly affect the health and production capacity of cows in the lactation period.Knowing the body condition of cows in advance can increase the profitability of the farm and reduce unnecessary losses.This article proposes an automatic cow body condition scoring system to address the challenging task of manually evaluating the body conditions of cow herds in large-scale breeding farms,and the currently low automation level of machine vision-based cow body condition scoring.This article selects the entrance channel of the milking parlor as the research scene,and uses the oblique upper view of the cow as the research object,and employs deep learning object detection algorithm as the technical support to design and validate an automatic approach for obtaining the individual body condition of dairy cows from video.We have completed the development of an integrated equipment for BCS evaluation and individual cow identification,as well as the construction of a comprehensive information service platform for the BCS of cows.This provides new solutions and technical support for future large-scale breeding farms.The main work and conclusions are as follows:(1)A deep learning based cow condition scoring model was constructed.A data acquisition device was built to capture images of the cow hindquarters in the milking parlour,and a cow condition dataset for training the model was created.The YOLOv5s model was used as the baseline,and the DPN network structure and CBAM attention mechanism were introduced into the original YOLOv5s network model to enhance the model’s ability to extract features related to cow body condition.By replacing the CBS module at the original network Head end with depthwise separable convolution,the model parameters were reduced while also improving the overall scoring performance of the model.Experimental results showed that the YOLOv5s+DW-CDPN model proposed in this article has high accuracy and low floating-point operation rate and the model shows precision,recall and average accuracy values of 94.3%,92.5% and 91.8% respectively on the individual cow image scoring tests,which increased by 3.1%,2.7% and 4.2% compared to the original YOLOv5 s network.(2)Aiming at the problem of individual separation and localization of cows in the video,an automatic detection positioning and segmentation algorithm for the hindquarters of cows in video frames based on spatiotemporal prior video frames is proposed,which is used to accurately obtain the body condition score of cows.First,the cow’s hindquarters are automatically detected by the improved YOLOv5s+DW-CDPN model;then the individual segmentation of each cow in the video is realized based on the spatio-temporal prior and coordinate tracking;finally,the best scoring result of each cow in the video is obtained.The algorithm can effectively achieve individual segmentation and accurate body condition scoring of a group of cows in a relatively complex environment in the milking parlor.In 10 complete video tests,the algorithm achieved 94.2% overall cow detection accuracy,92.5%body condition scoring accuracy among the accurately detected cows,and 87.1% overall cow body condition scoring accuracy in the video.(3)To implement and promote the proposed method in practice,the development of automatic cow body condition scoring equipment was completed.Firstly,the deployment of the cow body condition scoring model using Deep Stream and Tensor RT was completed.Secondly,the Jetson Xavier NX master control device,webcam,UHF RFID reading and writing unit and WIFI module were used to complete the corresponding embedded development,and the integration of each cow’s body condition and ID information was achieved through the correlation of the UHF RFID reading/writing unit and the webcam.Finally,the results were sent to the cloud server and data visualization was achieved through the construction of a comprehensive service platform for cow body conditions to provide farmers with intuitive information on the condition of cows.In summary,this study built a data collection device,completed the construction of a cow body condition scoring model,and realised the automatic segmentation of cow hindquarters and automatic body condition scoring for cows in videos.The model validation and development of the cow body condition scoring and individual identification,as well as the cow body condition information service platform,were designed and completed,providing an important technical solution for scoring the body condition of group cows in large-scale dairy farms.