Automated livestock monitoring is a promising solution for vast and isolated farmlands or cattle stations. The advancement in sensor technology and the rise of unmanned systems have paved the way for the automated systems. In this work, we propose an Unmanned Ground Vehicle (UGV) based livestock detection-counting system for fusion images using restricted supervised learning technique. For image fusion, we propose Dual-scale image Decomposition based Fusion technique (DDF) that fuses visible and thermal images. To reduce the difficulty of ground truth annotation, we introduce Seed Labels focused Object Detector (SLOD) that carefully propagates the annotation to all the object instances in the training images. Further, we propose a novel Restricted Supervised Learning (RSL) technique that produces competitive results with minimal training data. Experimental results show that the proposed RSL is more efficient and accurate when compared to other learning techniques (fully and weakly supervised). On the test data, with only five training images and five seed labels, the restricted supervised learning has improved the average precision from 4.05% (using fully supervised learning) to 80.58% (using restricted supervised learning). With 50 seed labels, the average precision is further boosted to 91.56%. The proposed model is extensively tested on benchmark animal datasets and has achieved an average accuracy of 98.3%.

自动化的牲畜监测是广阔而偏远的农田或养牛场的有希望的解决方案。传感器技术的进步和无人系统的兴起为自动化系统铺平了道路。在这项工作中，我们提出了一种基于无人机的基于受限监督学习技术的融合图像的牲畜检测计数系统。对于图像融合，我们提出了基于双尺度图像分解的融合技术（DDF），该技术融合了可见图像和热图像。为了降低地面实况注释的难度，我们引入了聚焦于种子标签的对象检测器（SLOD），该对象检测器将注释小心地传播到训练图像中的所有对象实例。此外，我们提出了一种新颖的受限监督学习（RSL）技术，只需最少的培训数据即可产生竞争性结果。实验结果表明，与其他学习技术（完全和弱监督）相比，提出的RSL更加有效和准确。在测试数据上，只有五个训练图像和五个种子标签，受限监督学习将平均精度从4.05％（使用完全监督学习）提高到了80.58％（使用受限监督学习）。带有50个种子标签，平均精度进一步提高到91.56％。所提出的模型已在基准动物数据集上进行了广泛测试，并已达到98.3％的平均准确度。