Towards efficient search methods in object tracking: An evaluation and application to precise tracking

Abstract

Object tracking is a much researched subject in the computer vision community. With more and more tracking algorithms reported every year, standard benchmarking and evaluation methods are reported for long term tracking systems. We present a public dataset to evaluate trackers used for human and robot manipulation tasks. For these tasks high degrees of freedom (DOF) motion of the object is to be tracked with high accuracy. Both the process of recording the sequences and how ground truth data was generated for the videos is described in detail. As an initial example, the performance of seven published trackers are evaluated. We describe a new evaluation metric to test sensitivity of trackers to speed. A total of 100 annotated and tagged sequences are reported. All the videos, ground truth data, tagged image frames, original implementation of trackers and evaluation scripts are made publicly available. We also introduce a new search method in tracking. Sequential Graph based Approximate Nearest Neighbour Search algorithm or SGANNS. It uses overlapping image features in videos to build a connected graph, offline. This graph is then searched efficiently during tracking to predict the best warp parameters. We test this algorithm on the dataset reported and further analyze the results. Finally we show that using a detection module, registration based trackers can be made more robust. We address tracking challenges of occlusion and varying appearance which a regular registration based tracker fails to track.