Tracking Ants Through Occlusions

The automated tracking of social insects, such as ants, could greatly increase the speed and reliability of gathering data used in studying their complex behaviors. Recently, data association based multipleobject tracking approaches have shown promise in improving tracking during occlusion. However, due to the high density and occlusion rate present in ant video, tracking is still prone to (1) identity switches with nearby ants coming in and out of long occlusions and (2) long periods of high occlusion where detection is difficult. We propose extending a previous data association based tracking method to address these issues. First, we introduce occlusion tunnels to constrain the possible choices allowed during the association process. Second, we propose a optical flow clustering method to detect and track ants during periods of known occlusion (or gaps) within the associated tracklets. Testing results on a 5,000 frame video of 50 ants shows that the inclusion of occlusion tunnels reduces tracking errors by 18% while our tracking with optical flow clustering reduces the number of missing detections by 59%.