Research On Object Tracking Algorithm Based On Correlation Filter

Object tracking is one of the most important research topics in pattern recognition and computer vision.It has a wide range of applications in the fields of video surveillance,driverless and human-computer interaction.Its technical aspects involve machine learning,deep learning and computer image processing,and have broad application prospects and research significance.After decades of research and development,a large number of excellent algorithms have been proposed successively,which greatly improved the accuracy and robustness of the tracking performance.However,in practice,since the target itself and the tracking background are constantly changing during the tracking process,factors such as target occlusion,fast motion,and illumination variations will all interfere the tracking results.Therefore,it is still very valuable in object tracking research.In order to further improve the accuracy and robustness of the tracking algorithms,this dissertation conducts in-depth analysis and research from the perspective of the tracking framework designing and the target feature selecting which is based on discriminative correlation filters?DCF?.The main contributions of this dissertation are summarized as follows:?1?Deep Features with Binary Matrix Mask Based Correlation Filter Tracking Algorithm?MDCF?is proposed.Based on the classical correlation filter tracking algorithm,the learned binary matrix is proposed as a local constrained mask to achieve pruning the template information and improving the tracking accuracy and robustness.The mask concentrates the template information and effectively alleviates the boundary effects caused by circular shifted training samples.Moreover,ResNet50 deep features are introduced in the process of feature extraction to replace HOG and CN features.When extracting deep features,the training sample set is expanded by exploiting rotation,flipping,and Gaussian blur operations to enhance expressive ability of the target template.Therefore,the tracking accuracy and robustness of the algorithm are further improved.?2?A tracking algorithm via Temporal Consistency and Spatial Pruning Based Correlation Filter with Multiple features?TCSP?is proposed.Based on MDCF,the ₂-norm is used as the temporal consistency model to establish constraints on the filters from two consecutive frames,so that the filter templates can learn the target context information and make the current filter template information and historical template information as consistent as possible.Temporal consistency model can mitigate the problem of filter degeneration caused by training samples contaminated and increase the anti-interference ability of the algorithm.The proposed algorithm is optimized by Augmented Lagrange Method?ALM?with fixed iterations.Comparing with the filter obtained by the traditional method,the filter template obtained by proposed method exhibits low rank characteristics.In addition,the proposed algorithm utilizes the structured color histogram feature as an independent template model to obtain the target response,which increases the reliability of the tracking result.Experimental results on standard tracking benchmarks show that the proposed algorithm achieves outstanding performance when dealing with occlusion,illumination variations and background clutter and etc.?3?The selection of target feature types affects the tracking results of the algorithm to some extent.To further improve the accuracy and robustness of the tracking results,under the proposed TCSP algorithm proposed in Chapter 4,this dissertation explores and compares the impact of six traditional features and four deep features on tracking results.Moreover,multiple features are combined and tested on standard benchmarks.The tracking results of multiple features are also compared,which further improve the accuracy of the algorithm.