Pigs’ Behavior Detection And Analysis Based On Video Tracking And Multi-modeling Audio Recognition

Farming industry is an important industry related to the livestock of people, improving the efficiency and environment of farming and achieving welfare aquaculture is the development trend of pig breeding. Making use of the intelligent detection method instead of artificial observation to detect and evaluate the health of pig is one of the foci of research. After the analysis of domestic and foreign research of the object tracking and sound recognition, based on the video tracking and various models of sounds recognition, the detection and analysis method of pig behavior is proposed in this paper. The processing technology of the digital image and the recognition technology of the sounds are used to reach the aim of pig behavior detection, like feeding, sleeping, moving, excretion and so on. At last, the evaluation and gradation of the detection results are carried out. The principal contents of the paper are as follows.(1) The object tracking technology of pigs based on video includes two aspects, objection detect and tracking of pigs. Aiming at the problems of detection holes and object missing when the object moves slowly or in stillness by GMM, this paper proposes an object detection method combined with GMM and Mean Shift based on the introduction and analysis on the common object detection methods and specific characteristics of pigs. After segmenting by Mean Shift, the Two-Pass is used to mark the objects. It is possible to obtain complete object contour by combining the detection results of two methods. In order to solve the problem of trajectory drawing cross and object tracking cross in the process of object trajectory tracking, the object detection method of pig combined with particle filter and frame by frame detection is proposed. As the camera lens’ focal distance is unchangeable, the area of object is essentially constant during video capturing process. The object location can be demarcated by the minimum rectangle and then calculating the area of the rectangle window. It shows that the detect result may be wrong when two pigs contact if the area is greater than the threshold. At this time, detect the number of objects, if the number is one, the method of segmentation by pixels’ threshold is used to segment object at this time. Otherwise using particle filter to predict the next position of the object. After the comparison of object positions between the predicted result and next frame in the video, Drawing the trajectory of pigs according to connecting the positions of object by the improved nearest neighbors. Finally, a visual interface is made for direct operating and displaying by MFC.(2) Through the analysis of the sounds of pig in different states, the models are made to recognize pigs’ sounds which are in eight kinds of states. Aiming at the problem of poor accuracy of sound recognition by single model, this paper proposes a new method of the optimization of multi-model’s results. Firstly, the feature parameters of pig’s sound under different states were preprocessed after de-noising, endpoint detection and adding windows and framing. Then train the SVM, HMM, and Adaboost by optimized feature parameters. Finally, the three trained models were used to recognize pig’s sounds and then optimize the best result by the method of feature matching as output. The experimental results show that the correct recognition rate was improved after optimization of multi-model’s results than single model.(3) According to the detection results of video and audio, behaviorally anchored rating scale is used to establish different scoring systems. Then scoring the behaviors of pig. Fuzzy weight distribution rule is used to making weighted summation of different behaviors’ scores.At last, making reference opinions about the pig’s health condition by the score.