Research On Video Object Tracking

As one of the fundamental problems in image and video processing,computer vi-sion,and pattern recognition,video object tracking is crucial to application in numerous fields such as military defense,public security,smart city,intelligence transportation,to name a few.Video tracking has broad application prospects and important theoretical research value.Although much progress has been made in the past decades,developing a robust and efficient video object tracking algorithm is still a challenging problem due to the complexity in intrinsic factors(pose variation,rotation,deformation,etc.),and extrinsic factors(illumination variation,scale change,occlusion,background clutter,etc.).In this thesis,we conduct research towards robust and efficient video tracking through three different perspectives,leading to better video tracking performance.The major contributions of this thesis are as follows:1.Proposal of a video tracking algorithm based on level set segmentationVideo tracking can be achieved via frame-by-frame image segmentation.How-ever,there exist problems such as inconsistency of color or illumination brightness in video frame when using traditional intensity-based segmentation methods.To this end,a video tracking algorithm is proposed based on level set segmentation to embed 2D object contour into 3D surface.We achieve curve evolution indirectly via surface evo-lution,and drive the zero level set to the target boundary by several evolutionary forces to achieve object segmentation.We enhance the level set segmentation by introduc-ing the steerable filter to discriminate the gradient image.In this thesis,experimental evaluation on VOT-2018 is carried out,and the method is applied to medical image rib tracking and respiratory state analysis,which verify the effectiveness of this method.2.Proposal of a visual tracking algorithm based on residual network and correlation filterThe time cost of the tracking method based on correlation filter mainly comes from two aspects:the feature extraction process and filter construction process.It is redun-dant for the artificially designed features(e.g.,HOG and Color Name)of video tracking tracker based on real-time correlation filter.To address this issue,a lightweight off-line learning method is designed to compress the visual features,and the correlation filter is used as a layer in the network to adapt the learning model to the tracking task.In addition,the residual structure is used in the training process to avoid the problem of overfitting.The network model presented in this thesis is simple and effective.The experimental results show that the proposed method can be applied to the existing cor-relation filter based tracking methods and improve accuracy and efficiency.3.Proposal of a visual tracking algorithm based on multiple tracker fusionIn video tracking,universal tracker for various circumstances is lacking due to the diverse illumination changes,occlusion and target deformation.To cope with different challenging factors,a fusion framework is proposed to absorb the strength of differ-ent tracking algorithms for robust object tracking.The proposed framework measures the pair-wise correlation between different tracker pairs based on their appearance and geometric consistency.It identifies the unreliable trackers by analyzing the computed pair-wise relationships with two effective strategies,discards the potential failure track-ers and fuses the rest trackers in a weighting manner.The proposed approach is focused on the output fusion of different trackers regardless of their specific details,making our framework compatible with any new tracker.We demonstrate its effectiveness by ex-tensive experimental results on the challenging OTB-2013 and OTB-2015 datasets.