Research On Robust Tracking Algorithm For Complex Scenes

Visual tracking is of great significance to computer vision,and has been active in areas such as intelligent transportation,autonomous driving,and smart homes for decades.Visual tracking means that after a target is given in the first frame of a video sequence,the tracker continuously locates the target in subsequent frames and establishes a target moving track.In real tasks,in the face of complex scenes such as occlusion,scale changes,and motion blur,the tracker is very easy to lose the target.At present,Visual tracking is divided into traditional Visual tracking algorithms and Visual tracking algorithms based on deep learning.The advantage of traditional Visual tracking algorithm is that it has good real-time performance,but it designs feature extraction methods for specific scenes,which limits its generalization in different scenes.Visual tracking algorithms based on deep learning rely on the powerful feature extraction capabilities of deep neural networks and good generalization.They can achieve better real-time tracking results in different scenarios,but they have the problem of insufficient real-time performance.In addition,in the scene of long-term Visual tracking,their tracking effects are not stable.To this end,based on the similarity measurement tracking algorithm,this thesis conducts research from the aspects of obtaining robust and rich target features,adapting to changes in target motion scenes with high confidence,and so on:1.According to the tracking algorithm based on similarity measurement,the tracking accuracy and real-time performance are high,but there are characteristics of tracking drift and poor scale adaptability in long-term tracking scenarios.The research analyzes the feature extraction network and tracking strategy of the baseline algorithm SiamFC,and proposes a twin network tracking algorithm SiamFC-22 that combines feature fusion and dual template nested update under the premise of ensuring real-time performance.The deep residual network ResNet-22 deep network is used for target feature extraction,and the semantic response is constructed based on the deep features of its strong target recognition ability.Based on the shallow high-resolution features of ResNet-22,a specific structural position response with strong positioning capability is constructed;in the tracking strategy,the two responses are weighted and merged and the dual template nested update mechanism is used to update the template.After OTB2015 and VOT2016 data set testing,tracking is more adaptable to scenes such as fast movement and occlusion,and the tracking speed is 32 frames per second,which meets real-time requirements.2.So as to improve the scale adaptability of SiamFC-22 and reduce the amount of tracking calculation,a tracking algorithm UPSiamFC based on Anchor-Free is proposed.In order to strike a balance between the amount of calculation and feature extraction capabilities,the algorithm uses feature fusion and attention mechanisms to design and propose a convolutional neural network AlexNet-UP as a feature extraction network;in order to improve the scale adaptability of the tracker,use Anchor-Free regression network thought designed a regression response to predict the coordinates of the tracking frame;in the tracking strategy,the template update method based on the estimation of the tracking frame category is used to update the template.After the OTB2015 and VOT2018 data set tests,the tracker has achieved good scale adaptability,and the algorithm has reached a speed of 84 frames per second.3.In order to verify the long-term stable tracking effect of UPSiamFC,the long-term video sequence data set UAV20 L is selected for testing.UPSiamFC achieved an accuracy of 0.662 and a success rate of 0.502 in the experiment,and it can still track steadily even in long-term complex scenarios.In order to verify the tracking effect of UPSiamFC in actual engineering scenarios,a laboratory data set was used for testing on the TX2 platform.The experimental results show that UPSiam FC can track the flight target in real time at an average speed of 46 frames per second on the TX2 platform,which meets the application requirements of actual engineering.In summary,in response to the robust adaptation and high-speed real-time requirements of Visual tracking in complex scenes,we have carried out in-depth development from more robust semantic features,higher resolution concrete features,adaptive tracking confidence evaluation,and more accurate scale prediction.Research and propose the Visual tracking algorithm of SiamFC-22 based on residual network and the Visual tracking algorithm of UPSiamFC based on Anchor-Free.Through the experimental verification of public data sets and actual projects,it is shown that the proposed algorithm significantly improves the target in complex scenarios.Stable tracking ability and good real-time performance.