Multiple object tracking is a fundamental problem within computer vision and has a wide range of applications. Although well studied, it remains a difficult task especially in scenarios which contain many occluding and highly similar appearing objects such as in videos of social insects. Data association based tracking methods have recently become popular due to improvements in object detection methods. Rather than tracking sequentially, detections found independently in each frame are first associated in short trackings or tracklets across adjacent frames which are further associated into longer trackings. A key component of this tracking method is the affinity model which measures the likelihood that two tracklets belong to the same object and incorporates appearance, motion and temporal information. First, we propose to improve the affinity model within insect videos by introducing a new set of irregular motion features. Second, we propose a method for filtering nearby confusing associations by using a sequence of foreground blobs called "Occlusion Sub-Tunnels." Finally, we propose a method for online learning of the full affinity model by automatically generating a set of weakly labeled samples using connectivity within occlusion sub-tunnels.