Study On3D Data Capturing And Processing Methods For Live Animal

Observing, Learning and novel application of morphological structure of biology is a fundamental capability that enables us to establish a variety of application system in human lives. Geometric morphological traits of animal or plant are one basic information source for the precision agriculture in the domain of agricultural application system. At present, missing of approach of acquiring and processing geometric morphological traits of middle scale domestic animal has become a bottleneck for large scale livestock farming. Therefore, this paper explores3D data acquiring technology and post processing key methods for the middle scale animal body surface based on range camera in order to extract the geometric morphological traits continuously, high throughput and automatically in the future.In this thesis, we employed live pigs and humans as the experimental subjects, and proposed the3D data acquiring solution for the middle scale animal body surface based on commodity depth camera, and investigated the key methods of posting processing3D data of animal body surface, such as registration, segmentation and non-rigid3d shape matching. We also implemented the full view3D data acquisition application system which includes two depth camera based on the above research. More details as following:(1) For the first time we assess the feasibility and operating condition of commodity depth camera as3D data acquiring equipment for the animal body surface. For this purpose, the study selects Xtion Pro for data acquiring device, and chooses dairy cow model and dairy cow as the experimental subjects. In indoor environment, three dimensional laser scanner with high precision was employed to get point cloud data of cow model as comparison data and Xtion Pro was used for acquiring point cloud data in different distance away from cow model as test data. Statistical errors between comparison data and test data were calculated by the professional point cloud software to quantitatively analyze change laws of precision and density of point cloud data from Xtion with different distance in order to determine proper acquiring distance. Under the dairy cow breeding environment, body measurements were first determined manually. Then we used Xtion Pro to acquire point cloud data in the distance of less than1.2meters. Visualization of the point cloud data and interactively measurement of the cow body based on point cloud data from Xtion Pro were performed so as to qualitatively analyze the sunlight and surface materials’ influence on quality of point cloud data acquiring from Xtion under livestock farm condition. After that interactively measurement and manually measurement are comparatively analyzed. The test results indicate that both the average error and the relative error are less than±5mm and10%separately on the condition that the distance from Xtion to subject is between0.6m and1.2m. The operating conditions are determined and the Xtion Pro is appropriated for3D data acquiring equipment for the animal body surface.(2) An improved algorithm for registering3D point cloud data was proposed, which was based on combining newly introduced similarity metrics with sample consensus strategy for filtering wrong correspondence. Firstly we matched the local geometrical features encoded by the normal and curvature respectively in order to get the initial correspondences. Unlike existing Euclidean distance constraint based wrong correspondence rejecting approaches which can’t work properly on the3D scene containing the relative small non-rigid part, our random sample consensus correspondence rejecting strategy can handle both the scene of animal movement and static scene. Then the transformation aligning the filtered initial correspondences was estimated by using the singular value decomposition method. Both non-rigid synthetic model and3D sequences of real world animal moving were used to test our algorithm. The results show that the proposed algorithm exhibits faster convergence than the original method and can be used for non-rigid model initial alignment, camera motion estimation.(3) A novel approach of rigid or non-rigid object tracking and segmentation from3D range image stream was presented in order to serve the purpose of segmenting the animal from3D scene. This method began by initial aligning the two adjacent key frames in order to estimate the relative camera pose of two frames. Then the moving objects were detected by frame difference method combined with view constraint so as to avoid the influence from background movement. Ray tracing method was used for eliminating the false moving object caused by occlusion. The two adjacent key frames were initially clustered in space domain based on support part removing. Meanwhile the correspondences based on geometrical feature matching and closest criterion between them were found in order to construct the matrix which integrated the information of topology of space domain and consecutiveness of time domain. The matrix can be used to localize (spatially and/or temporally) regions where segmentation results are over segmentation or under segmentation; or combined to yield a corresponding cluster to indicate the tracking results if the segmentation results are good. These clusters from initial segmentation were tracked, split and merged over the entire sequence to produce the final results. We showed experiments in range image sequences in which multiple people or animals come in close proximity, enter and exit the scene. Qualitative experiments using many real life datasets show that our algorithm effectively handles motion segmentation and tracking of complex scenes such as non-rigid motions, severe occlusions, multi-object, camera movement and moving object connected with each other or the background.(4)An approach for estimating correspondences between two non-rigid shapes that can handle articulation and deformation of the surfaces to be matched was proposed and potentially useful for a number of applications, including non-rigid3D reconstruction, pose normalization of3D models of animal. Under the assumption that limbs are pitifully thin compared to the rest of bulk for the common animals, the method is capable of automatically discovering the articulated parts of the surface without requiring knowledge of the topology or the number of rigid parts. Processing began by estimating potential sparse correspondences between the source and the target surface. These were used to align the largest corresponding parts of the two surfaces. Fragments of the surface that are not consistent with this alignment generate part hypotheses on which the algorithm is applied recursively. The kernel correspondences were estimated by combining the geometrical feature matching and efficient spectral method on the largest corresponding parts. Then the kernel correspondences and connectivity of fragments with largest corresponding parts were used to estimate the alignment parameter of corresponding fragments. After the kernel correspondences for all parts had been estimated as above, we performed a final propagation step to compute dense correspondences for all points. We presented qualitative and quantitative results on three datasets comprising open and closed surfaces. The results show that our approach can handle large articulated motions and does not suffer from ambiguities due to symmetry, make less restrictive assumptions about the data representations, noise degree.(5) In order to overcome the limited field-of-view of singe depth camera, based on multiple depth cameras, a real time application system which can acquire3D data of animal body surface was proposed. An automatic ball-target-based offsite method for the extrinsic calibration of reconstruction system which is based on RANSAC was exploited. Easy adjustable depth camera fixing mechanism which can save the extrinsic parameter of depth camera temporarily was designed for fixing the depth camera and combined with ICP registration for the second calibration of depth camera onsite to avoid the error caused by resetting of system. The extrinsic parameters onsite can be used for registering point clouds acquired synchronously from depth camera in real time. And then, Interactive measuring method which is optimized in picking mode for body measurement is used to get shape traits of animal. Point clouds data acquired from high precision laser scanner is applied to evaluation of the calibration method with the pig specimen as test subject. Live animal and people were used to test the function of depth camera fixing mechanism and body measurement error of live subjects. The results show that the ball-target-based method can get the extrinsic parameters of depth camera automatically. The speed of reconstructing the whole pig can reach in excess of15fps. Body measurement errors of both static and dynamic subjects are less than4%and6%respectively. The depth camera fixing mechanism can be adjusted for different size animal and effectively avoid the calibration onsite like traditional multi-view system so as to eliminate the animal stress. So the system can be applied to body measurements in agricultural field.Finally, we summarized our conclusion and gave some suggestions about the3D data acquiring technology and post processing key methods for the animal body surface based on range camera for the future research.