Hierarchical Association And CNN For Online Multi-object Tracking

Tracking of multiple objects is an important topic with many computer vision applications such as robot navigation,human computer interactions,sports analysis,and video surveillance.In particular,as the use of video camera grows explosively,it becomes increasingly important to develop the method of robustly tracking multiple objects,especially people,in real-time.Due to the significant improvements in object detectors,a lot of recent works on multi-object tracking have focused on the tracking-by-detection strategy and much progress has been made.However,in complex scenes,it is still challenging due to frequent and prolonged occlusions,abrupt motion change of objects,unreliable detections,similar appearances of different objects,and so on.In this paper,we tackle two key aspects of multiple target tracking problem: implementing an efficient and accurate data association algorithm and designing an accurate affinity measure to associate detections.As for the first contribution,tracklets have been divided into two parts by computing their confidences.Tracklets with high confidence are locally associated with online-provided detections,while tracklets with low confidence,which are more likely to be fragmented,are globally associated with other tracklets and detections.To be more concrete,the tracklets with high confidence are first considered to be locally associated with detections.For the second contribution,the original pair of input patches can be changed into four patches by a central-surround two-stream network and we simply consider the four patches as a multi-channel image,which is directly fed to the first convolutional layer of one convolutional neural network(CNN).The network are pre-trained on the auxiliary data,and then we combine online transfer learning for improving appearance discriminability by adapting the pre-trained deep model during online tracking.We combine the discriminative appearance model and the non-liner motion model as the affinity measure.Our method shows encouraging results on many standard benchmark sequences and significantly outperforms robustly in crowded scenes with long-term partial occlusions.