Piglet tracking is critical to automated piglet behaviour and welfare analysis. Following the tracking-by-detection paradigm, we have developed an online piglet tracking network (OPTN) composed of a base network, a detection head and an association head. The unweaned piglets in video images were detected at the region-based detection head, and the piglet central locations were mapped onto the feature maps produced by the base network and extended network to form central feature vectors. Then the exhaustive central feature vector permutations were input to the affinity estimation network to generate an affinity matrix that accounted for the affinity between the video image pair detections. To make full use of the tracking history, we used the Hungarian algorithm to optimise affinity prediction with affinity accumulation and designed a distance-based tracking state adjustment strategy to correct false state prediction and recovered lost IDs. Our method achieved favourable tracking performance with an IDF1 score and MOTA of 96.55% and 97.04%, respectively, and an inference frame rate of 6.89 fps. Our method outperformed popular MOT methods, such as SORT, SST and CenterTrack, in short video clips and a long video episode. OPTN was robust against large illumination variations with a video rate as low as 1 fps. Our computer vision-based piglet tracking method may aid animal tracking-related behaviour analysis as well as piglet surveillance.

仔猪跟踪对于自动仔猪行为和福利分析至关重要。遵循“按检测跟踪”范例，我们开发了一个在线小猪跟踪网络（OPTN），该网络由一个基本网络，一个检测头和一个关联头组成。在基于区域的检测头处检测视频图像中未断奶的仔猪，并将仔猪的中心位置映射到由基础网络和扩展网络生成的特征图上，以形成中心特征向量。然后将详尽的中心特征向量置换输入到亲和力估计网络生成考虑视频图像对检测之间的亲和度的亲和度矩阵。为了充分利用跟踪历史，我们使用了匈牙利算法来优化具有亲和力累积的亲和力预测，并设计了基于距离的跟踪状态调整策略来纠正错误状态预测和恢复的丢失ID。我们的方法获得了良好的跟踪性能，IDF1得分和MOTA分别为96.55％和97.04％，推理帧率为6.89 fps。在短视频片段和长视频片段中，我们的方法胜过SORT，SST和CenterTrack等流行的MOT方法。OPTN具有强大的鲁棒性，可应对低至1 fps的视频速率的较大照明变化。