Research On Object Detection And Part Boxes Combination Technology Of Facility Cow Based On Deep Learning

Now,the artificial intelligence,big data and robot technology are developing rapidly,and the inspection robots are widely used in intelligent breeding.Object detection can provide the category and positioning information of livestock and poultry objects for inspection tasks.However,the problems such as complex breeding environment and facilities,serious objects occlusion and changeable position and posture of livestock and poultry increases challenges for the object detection process.In the meantime,in the process of object detection,the detection of livestock and poultry individuals can get the overall information of them,and the detection of each part can obtain more specific local information of the object.However,the incorrect correlation between the overall information and local information will lead to the problem of object identification and positioning errors in the inspection process.Taking the facility breeding cows in RGB and infrared images as the specific object,the problem of target detection is analyzed.Because the traditional detection algorithm can not complete real-time and accurate object detection,the object detection algorithm of deep learning is selected to complete the task of cow object detection.Firstly,RGB and infrared cow data sets are constructed,and four kinds of fast RCNN,SSD,Yolo v3 and Yolo v4 are selected from the deep learning algorithms.The two cow data sets are compared respectively.Finally,Yolo v4 with the best performance is selected as the basic algorithm of object detection.In view of the fact that the above-mentioned actual breeding conditions directly affect the object detection accuracy of cows,the Res PSA-Yolo v4 algorithm is proposed,that is,the Res PSA module combining residual unit and attention mechanism is added to Yolo v4 network to enhance multi-scale semantic information and object area features,and avoid network degradation while broadening the network depth.At the same time,our self built data set are clustered by the algorithm of Kmeans++ in order to get the best anchor template.Using the improved Yolo v4 algorithm,the detected m AP on RGB and infrared cow data sets reach 77.99% and78.29% respectively,which are increased by 3.65% and 4.54% respectively compared with the original Yolo v4,and the recall and AP of each category are improved,also ensuring the real-time detection.Aiming at the whole cow and the object detection output of head and tail,a object position combination algorithm is designed to realize the association combination of the whole Bounding Boxes and head and tail position Bounding Boxes of the same cow.Firstly,the preliminary position combination is carried out based on the Io P（Intersection over Part box）threshold judgment,and then the "elimination method" and the "one to several" and "several to one" position recombination methods based on Mask RCNN are used to solve the wrong combination caused by dense cow objects.Based on the data of experiment,the position combination accuracy of RGB and infrared object bounding boxes can reach98.20% and 97.89% separately.