Sow Posture Recognition And Analysis Based On Depth Image

Visual perception of pig behavior has become a hot topic in the field of precisionlivestock breeding.The posture of lactating sow can provide basic information for the study of maternal behavior characteristicsand regularity,which is an important index of sow maternal behavior evaluation.However it is laborious and subjective to observe its posture by artificialor video.The study of automatic real-time recognition technology of sow posture based on computer vision not only reduces the labor intensity and breeding cost of the staff,but also timely finds the abnormal posture of the sow and takes corresponding measures.However,due to the influence of external illumination and shadows on the RGB image,this paper takes the K inect sensor to obtain t he depth image,establishessow posture depth image library,and studies sow detection algorithm based on the depth image,and based on the target area obtained by the detection,the automatic identification algorithm of sowposture based on deep convolution network is studied.The work of this paper mainly includes the following aspects:Establishment of sow posture depth image library.Based on the analysis of lactating sow posture,the sow depth image was collected with the K inect sensor vertically downward.Due to the noise problem in sow depth image,the Median F ilter is used to remove the image noise while maintaining the detail information o f the image.Then,the Contrast Limited Adaptive Histgram Equalizationis used to improve the contrast between the sow and the background in the image,so as to obtain higher quality sow depth image.Then,according to the behavior characteristics of sows,five different postures are defined.Finally,the classification of the sows posture is manually labeled,which constitutes the sample library for the detection and recognition of the sows.Research on sow detection algorithm based on depth image.The change of body appearance of sows with different postures,which brings great difficulty to target detection,the sow detection method of the Multi Deformable Part Model is studied.A model of multideformable parts with optimal sow detection was obtained by repeated adjustment of model structure.The model integrates four different deformable part models,each of which contains seven component filter models.Then,by optimizing the model,the average detection time is reduced by 5 times while maintaining the detection accuracy.Finally,it is proved that the method can detect the sow in the image better,the accuracy rate is 93.3%,which is 18% higher than the detection algorithm based on HOG + SVM.Research on the algorithm of sowposture recognition and analysis.The AlexNet convolution neural network is used to adjust the network structure and optimize the parameters to adapt to the recognition of the sows in the piggery scene.The experimental results show that the network structure adjusted reached 97% on the accuracy of posturerecognition and indicate the effectiveness of the recognition model based on AlexNet,then through the comparison shows that the recognition performance is better than the traditional pattern recognition method.Finally,select a time segment with different posture of sows and analyze the posture of sows in this period,by comparing the results with the artificial statistical results,further illustrate the applicability of this recognition algorithm.