Target Tracking Technology Based On UAV Vision

As one of the most important research fields of computer vision,the object tracking has been widely used in smart transportation,human-computer interaction,biomedical and military fields.However,a large number of uncertain factors,such as: illumination variations,scale variations,deformation,occlusion,rotation,background clutters,etc.,affect the accuracy of the tracking algorithm seriously.The tracking with the unmanned aerial vehicle as the camera carrier is more difficult,because the target is far away from the camera,the target and background with inconspicuous features are captured.Tracking is even more impossible under the severe fog weather.In recent years,the kernelized correlation filter plays an important role in target tracking field due to its high speed and robustness.Two algorithms based on the kernelized correlation filter are proposed to solve above problems according to the research about the sensitivity of different features to the video attributes.The main work of this paper is as follows:（1）Sensitivity analysis of different features to the video attributes is carried out.After classifying the OTB-100 dataset according to the video attributes,the kernelized correlation filters with the Histogram of Oriented Gradient,the Color Names and grayscale are used to track these videos respectively.The accuracies of the three features under different scenes are compared.We can select the best features based on the experimental results to track something under specific scenarios in the future work.（2）In order to improve the tracking performance,the algorithm integrated HSV color space and Histogram of Oriented Gradient is proposed.Three correlation filter trackers are trained by the Histogram of Oriented Gradient of hue image,saturation image and brightness image respectively.Then the prediction of the tracker with the biggest sidelobe peak ratio is chosen as the tracking result of this frame.Finally,three kernelized correlation filters are updated according to the tracking result.（3）For severe haze weather,the tracking algorithm cannot locate targets accurately.The tracker based on the dark channel defogging and kernelized correlation filter is proposed.First,we judge whether it is needed to defog according to the dark channel image,and the defogging is carried out if necessary.Then the object in the picture after defogging is tracked by the kernelized correlation filter.Finally,the trackers are updated according to the tracking results.In addition,the two tracking algorithms proposed in this paper are compared with existing tracking algorithms through extensive experiments.The results show that the proposed algorithm has higher accuracy and success rate compared with other trackers.The experimental results show that the integrated target tracking algorithm with multi-feature fusion is 3% higher than the kernelized correlation filter on the OTB-50 dataset,while the results on the UAV123₁0fps dataset show an 8.3% improvement over the kernelized correlation filtering algorithm.When the density of fog is high,the kernelized correlation filter algorithm cannot accurately locate the target.In this paper,the dark channel defogging tracking algorithm can still obtain better tracking results.