| Single object tracking is one of the important research topics in computer vision.It contributes to the convenience and intelligence of our daily life.With the development of hardware technology,the performance of tracking algorithms has improved considerably.And the practical application scenarios have been continuously expanded,including security systems,intelligent monitoring,unmanned aerial vehicles(UAVs)tracking,etc.Therefore,the reserch of single object tracking algorithms and their application is meaningful.It contains significant value for theoretical research and practical applications.However,object tracking has always faced several challenges.In general tracking scenarios,the main challenge is the variation in object appearance due to occlusion,motion,etc.,which can further lead to inaccurate similarity metrics and the tracking drift.Most existing tracking algorithms eliminate the negative impact of object appearance changes by updating the entire object template.However,when the object is partially occluded,the local information of the image is inaccurate and needs to be updated.In this case,updating the entire object template will conversely reduce the tracking precision.To address the above issues,the thesis proposes a tracker of reverse joint sparse representation with local template update.The tracker proposes a strategy for locally updating the object templates,thus reducing the effect of occlusion on the tracking results.And an rnverse joint sparse representation model is constructed to improve the tracking precision of the algorithm by reducing the reconstruction error between the candidate images and the object templates.Our method takes full advantage of temporal similarity and spatial continuity during tracking.It can be improve robustness and eliminate tracking drift compared with similar algorithms.The challenges in the UAV tracking scenario including:First,the small object scale makes it difficult to extract enough feature information.Second,the cluttered background makes it difficult to distinguish between background and object information.Third,interference such as occlusion and motion blur can frequently change the appearance of the object,leading to tracking drift.Last,due to the limited computating resources of the UAV,the computational effort of the algorithm must be strictly controlled.It will ensure that the algorithm is realtime.There are few existing UAV object tracking algorithms,and most of them still have much room for improvement in the processing of features.To address these issues,the thesis investigates a real-time UAV object tracking method based on the discriminative correlation filter.Our method introduces joint sparse regular terms as weight vectors in the correlation filtering model.By weighted fusion of multiple features,the algorithm selectively improves the ability to extract feature information.In addition,the group sparsity of multi-channels is used to construct spatial regular terms adaptively,and we introduce global response oscillations and local response variations into the update strategy.It enables the algorithm to correct errors and resume tracking,ensuring smooth consistency of the model and effectively eliminating tracking drift.The method successfully balances algorithm speed and accuracy with enhanced tracking robustness. |
