关键词: 奶牛, 肢体分解, 深度图像, 骨架提取, 形状视觉显著度, 分割线优先级

Abstract: The accurate movement details of cow obtained by cow limbs are the foundation of posture detection, behavioral analysis and understanding. In order to realize the accurate decomposition of head, neck, torso, forelimb, hind limb and tail，this paper studied and proposed a method for cows limb decomposition based on skeleton feature. The cow target was extracted from the depth image using depth information. A skeleton of cow target was produced using robust distance transform method, and the skeleton of redundant was pruned by contour partitioning with discrete curve evolution. The candidate split points, which were used to generate split lines, were retrieved by the junction points on the skeleton of cow. The shape visual saliency and split line priority were used to discard redundant split lines. The remaining split lines were selected to decompose the limbs of cow. The result showed that the average correct rate of cow each limbs was 95.09% and the correct rate of tail, which was most difficult to segment, was 9551%. Using the method of this paper to decompose the different body postures of cow, such as upward, normal walking, micro bow, and bow,  the average correct rates was 95.18%, 95.00%，94.85% and 96.23%, respectively. Thus, a high precision body decomposition of cow with different posture was realized.

Key words: cow, limbs decomposition, depth image, skeleton extraction, shape visual significance, split line priority