Research On Correlation Filter Based Air-to-Ground Object Tracking Algorithm

Visual object tracking is a fundamental and important issue in the computer vision.The research on object tracking technology is mainly to improve the robustness of tracking and speed up the running of the algorithm,but the two are usually contradictory.Object tracking is widely used in many fields,including intelligent monitoring,UAV detection,human-computer interaction,medical diagnosis and so on.In different application scenarios,the characteristics of the target in the image are different,thus the focus of the tracking algorithm is different.In the application of air-to-ground object tracking,there is a high demand for the speed of the tracking algorithm.And there is also a high demand for the anti-occlusion ability of the tracker,because the occlusion caused by the buildings and trees are the most common interference in air-to-ground object tracking.In recent years,correlation filter(CF)based tracker have developed rapidly.Both on the robustness or computational efficiency,the CF based tracker have achieved a state-of-the-art performance.In the article,we mainly employed two stateof-the-art CF based tracker,BACF(Learning Background-Aware Correlation Filters)tracker and CEO(Efficient Convolution Operators for Tracking)tracker,as our baseline tracker.And how to improve their computational efficiency and improve the antiocclusion ability when using them for air-ground object tracking is studied.Firstly,we proposed and experimentally verified the mutual constraint relationship between the sampling intervals of tracker's training,running speed of tracker and video frame rate.At the same time,it was also verified that appropriately increasing the sampling interval of the tracker's training can not only improve the tracker's average operation speed,but also reduce the risk of overfitting.Secondly,two algorithms for improving the anti-occlusion ability of the CF based tracker were proposed.They are region color histogram based algorithm and image segmentation based algorithm.The core ideas of the two algorithms are identical and the method includes two steps: step 1,perceive occlusion and evaluate occlusion levels;step 2,Based on the occlusion evaluation result,the tracker chooses to enter two different working modes,normal mode and lost mode.When there is no occlusion or slight occlusion,the tracker is in normal mode.At that time,the search is performed normally and the training on the sample is performed by using an adaptive learning rate associated with the occlusion level.The tracker enters the lost mode when severe occlusion occurs or the target disappears.In this case,the sampling and training will be paused and the search area will be expanded to facilitate recapture of the target.These two algorithms effectively prevent the tracker from continuing to use bad samples for erroneous learning when the target is occluded,thus ensuring that the tracker remains aware of the target after temporarily losing the target.In this paper,extensive experiments have been carried out on the standard benchmark OTB50,OTB100 and VIVID,which proves that the CF trackers updated by our proposed algorithm achieved a significant performance improvement in anti-occlusion.