Research On The Behavior Identification Of Pig Delivery And Piglet Tracking In Video Surveillance

Detecting real time sow delivery and judging the movement state of birth piglets accurately by using intelligent monitoring instead of human observation are valuable and significant task in this field.And this topic is one of the main key technologies in a new product research and development of enterprise cooperation project.In this thesis,we propose the methods for recognizing the behaviors of pig delivery and tracking the piglets.The main research are summarized as below:1.Foreground detection.We study three commonly used foreground detection algorithms in computer vision:GMM(Gaussian Mixture Model),Vibe algorithm and inter-frame difference algorithm.After applying them to the real pig behavior videos,analyzing and comparing the real results,we propose an improved foreground detection algorithm in this thesis.And on this basis of improvement,the use of morphological operations in image processing helps to access better foreground results.2.Pig birth behavior recognition.We propose an algorithm for obtaining the location of sow and setting the region of interest adaptively based on image segmentation and the Hough transform,which successfully achieves the obtainment of sow location and the founding of the region of interest.Then,according to the behaviors of sows,we present a birth recognition algorithm based on the area of the region of interest and aspect ratio.On the real scene dataset provided by the enterprise partners,experiments display that the proposed methods are good for the location of the pig and the first piglet birth identification,and the accuracy is more than 95%.3.Piglet tracking.We first studies the piglet target extraction.Besides,we propose a fusion algorithm to obtain the complete outline of piglet.Then,we improve the traditional compression sensing algorithm,and propose a compression-aware piglet tracking algorithm based on the regional update of the region of interest.Experiments show that it can still be robust to track the piglet with our tracking algorithm even for some occlusion environments.Based on the above algorithms,we construct a prototype of pig delivery behavior recognition and tracking system,and and test it on the real scene data set provided by the enterprise.The results show that our system is suitable for the sow position and first piglet birth recognition,and the accuracy rate is more than 95%.Besides,for partially obscured scenes,it can still be robust to track the piglet.