Study Of Pig Skin Temperature And Gait Features Extraction Method Based On Multi-Source Images

Today, the gradually scale development trend in the pigs cultivation field has raised higher and higher demand to the automation, informatization and intelligentization of cultivation. As the computer vision based monitoring method for pig abnormal has the characteristics of non-contact and continuing automatic monitoring, it may help to decrease the loss of cultivation enterprise by timely detecting the pig diseases, early treatment, adopting quarantine measures. Moreover, it would help the farm staff to reduce the labor intensity, to improve the working environment, and to cut down the risk of human and animal cross infection caused by close contact with pigs. Therefore, the research of this field has important scientific value and broad application prospect.Current research in this field mainly use visible light images or videos to monitor and recognize pig activities. The features which can extact are limited. The accuracy of monitoring and recognizing is easily disturbed by illumination and environment. After broadly investigating and analysing, an multi-source image acquisition platform is constructed including visible light images, infrared thermographs, and depth images. In this dissertation, the work mainly focus on the automatic extraction method of pig features based on multi-source images, including the skin temperature feature of region of interesting (ROI) and the gait frequency feature. The contributions of this dissertation are summarized as follows:1. An auto registration method of infrared thermograph and visible light images of pig is proposed based on contour match of radial line feature points. Because of the different imaging principle between infrared thermal image and visible light image, the traditional image registration methods for visible images can not be suitable. In this dissertation, by analysing the characteristics of the two different source images acquired by thermal infrared imager timely, the image registration transform is simplified to an undetermined solving problem for the scale parameters between the two images. Furthermore, The feature space is constructed from the feature points set, which are the cross points of the auxiliary radial lines and the edge of contour. The weighted partial Hausdorff distance is used as the similarity measurement. The optimal scale parameters are obtained by iteratively finding the minimum distance between the feature points set of infrared thermal and visible light images. By using a RPROP algorithm to accelerate the iterative process, the auto registration of the two kinds of images is realized finally. Test results show that the accuracy of registration can meet the demand of further process of image fusion.2. A fusion method for visible light and infrared thermal images of pig is proposed,which is based on non-subsampled contourlet transform (NSCT). After image registration, to promote the ear area inspection results, the multi-scale and multi-directional decomposition are taken to the registered images by using NSCT firstly. Secondly, considering the effects of both the neighbor area energy and the co-variance on the image characteristic, a weighted fusion rule for the low-frequency coefficients is designed. The weight values are calculated according to the neighbor area energy and co-variance. And a rule of maximum neighbor area energy is adopted to fuse the band-pass coefficients. Finally, the fusion image is obtained by taking the invert non-subsampled contourlet transform to the fused coefficients. Tests show that the fused images have better vision effect and contour segmentation effect than visible light image or infrared thermograph.3. Aiming at the auto inspecting of the pig skin temperature of ROI with infrared thermograph, a series of experiments are carried and the pig ear root area is confirmed as the most meaningful ROI of skin temperature feature. Based on that, a pig ear area target inspection method with adapted Active Shape Model (ASM) algrithm is proposed firstly. On the basis of classic ASM, the NSCT decomposition coefficients obtained in the processing of image fusion are used to construct the local appearance models, and the skeleton extracting and matching method is using to improve the initialization of the average shape. With these adaptions, the pig ear area can be recognized effectly. Furtherly, the pig ear root area is drawn by offsetting the ear root contour. Finally, the feature of ear root skin temperature (ERST) is extracted by statistical analysis the relative area temperature data of infrared thermograph. Comparing the temperature statistical results obtained by the manual method and the proposal method, it is found that the proposal method is valid.4. A pig gait descriptor constructing method is proposed based on the analysis of depth image skeleton endpoints. Aiming to solve the problem of the computer vision description of pig gait, the depth image sequences of single pig walking are acquired by a depth camera from side view. Background substracting method is used to obtain the binary foreground images. Firstly, a series of processes, such as image skeleton extraction, pruning, and skeleton matching, are taken to locate the skeleton endpoints of four legs. Further, the depth values of neighbor skeleton points of each skeleton endpoints are drawn to determine the left or right property. Finally, the horizontal ordinate varying relation of the skeleton endpoints between fore-left and fore-right legs or hind-left and hind-right legs in the image coordinate space are used to. construct the pig gait descriptors and to set the pig gait description of computer vision-The proposed gait descriptor constructing method provides a foundation for analysing the pig gait feature with depth image sequence.5. A gait frequency detecting method is presented which is based on segmentation of gait descriptors sequence and stride reconstruction. The descriptor sequence extracted from the depth image sequence of pig walking is regarded as a single variable time sequence. Firstly, the descriptor sequence is segmented into five basic gait units, such as positive-uprise, negative-uprise, positive-downward, negative-downward, and stance by taking a twice segmentation. Then, orderly combining the basic gait units, the sequence can be transform into a combination of single step modes, including positive-cross step mode, negative-cross step mode, positive lameness step mode and negative lameness step mode. The relative step numbers of the sequence can be drawn out by counting the numbers of single steps in the combination. At last, the gait frequency of the sequence can be easily calculated by dividing the step numbers with time length of the sequence acquired. Experiment results show that the proposal method can extract the gait frequency features of pig walking accurately, even under the situation that there are the lameness behavior and long stance. The proposal method not only provide an effect way to evaluate fast or slow status of pig walk, but also may benefit to the study of pig lameness detection by further analysing the different combination of single step in a sequence.The works, of this dissertation have fairly academic meanings and applicable value on constructing and perfecting the computer vision based pig cultivation monitoring system, richening the monitoring features, and promoting the early warning ability.