UAV Visual Object Tracking Based On Keypoints

Visual object tracking is aimed to infer the location of the target in adjacent frames without prior information of the appearance of the object.Generic object tracking has made significant progress these years.However,there are only few literatures related to UAV tracking.UAV tracking is aimed to infer the location of the object from the videos captured by an aerial viewpoint.Compared to generic object tracking,the challenges mainly focus on fast motion,scale variation and aspect ratio variation.Generally,the object can be represented with some keypoints.The keypoint is able to discover local details and patterns of the object instead of trying to capture the semantic information of the whole region and is widely used in many fields like face recognition,pose estimation,gesture detection,etc.The local keypoint reflects the tree-structured structure of the object and connection of different parts.Therefore,by taking the advantage of powerful feature capturing capability of keypoints,this paper leverages keypoints to improve the performance of UAV tracking from the aspects of feature detection and search strategy.In tracking-by-detection framework,the traditional feature detector requires manual design for robust performance.The hand-crafted detector for feature extraction is usually applied to several specific objects of interest(e.g.,faces,humans)and thus not suitable for generic tracking problems where the object suffers from significant appearance changes,especially from an aerial viewpoint under UAV tracking scenario.In this paper,a sparse feature detector based on sparse coding is proposed to improve tracking robustness.The basic idea is to learn a dictionary from the samples in the previous frames and construct feature representations to represent the object during detection in the current frame.The samples are patches centered at the keypoints based on an adaptive FAST detector.The dictionary is learned with sparse coding for sparse representations and atoms of the dictionary are grouped to describe the local orientation of these samples.The final features are built after rectification of the sparse representations.The correlation filter is used to infer the object location.The qualitative and quantitative experimental results on UAV123 show the advantage of the proposed tracker against current state-of-the-art trackers in terms of accuracy.The general tracking-by-detection framework tends to sample patches locally around the center of the object location from the previous frame when searching for object at the current frame.In UAV tracking,the variation of camera viewpoint caused by UAV motion makes the local search strategy less reliable.Therefore,based on smooth optical flow,a search strategy beyond local sampling is proposed to provide object proposals for candidate detection under aerial scenarios.The main idea is to estimate the optimal smooth motion field by minimizing the objective function composed of three terms: a data term measuring the residual between the smooth optical flow and the original flow,a smoothness term enforcing the similarity of keypoint motions in the local motion field and a rigidness term preserving shape of object.First,a FAST detector with local threshold is employed to detect keypoints based on the intensity strength distribution in the local region.Then,the Lucas-Kanade flow is smoothed by solving the minimization with a Jacobi solver to infer the displacements of the keypoints.The object proposal is computed by estimating the translation and scale variation of the object.The results on UAV123 confirm the superiority of the proposed tracker in accuracy.The experiments are performed on UAV123 dataset which is obtained from the videos captured from aerial viewpoints.The evaluation is conducted under two metrics: precision and success rate.