Tracking with DAG Recurrent Neural Networks and Motion Revising
Cailing Wang, Zhenfei Zang, Huajun Liu, Xiao‐Yuan Jing · 2018
This paper presents a novel tracker which model the object's self-structure with DAG-RNNs and build the sequential dependencies. In contrast to existing trackers, our method innovatively utilized the undirected cyclic graphs(UCGs) to model two-dimensional image data, and decomposed the UCGs with directed acyclic graphs(DAGs), which convert the 2-D image data to be processed by recurrent neural networks. Through this method, we obviously improved the network's capability of distinguishing target from distractors. Besides, we use the motion revising layer which has 11 defined motion units and 2 classification units to capture the motion state among sequences and revise the position of bounding box to achieve a more accurate output. The experimental results demonstrate that our method achieved an outstanding performance, especially in several challenging aspects such as occlusion and motion blur.