Research On Correlation Filter Object Tracking Algorithm Based On RGBD Data

Visual tracking is an important building block of applications such as video surveillance and human-computer interaction.The object tracking algorithm determines the position and scale of the target in subsequent frames of the video sequence by the bounding box in the first frame of the video sequence,the bounding box is from human annotation or object detector.Object tracking algorithms based on correlation filters have achieved good performance and drawn widespread attentions.Correlation filtering uses a large number of cyclically shifted samples for learning,and transforms the correlation operations in the spatial domain to the dot-multiplication operations in the frequency domain,the complexity has dropped from(9)~3)to(9)7)2)9)).This thesis focuses on improving accurate of tracking algorithms,two improved methods based on RGBD are introduced.Firstly,a target aware discriminative correlation filtering object tracking method is introduced.RGB and depth images are used to calculate an accurate target background segmentation map for the search area,target background segmentation map is used to mask on the extracted features for reweighting.First,Color and depth histograms of target and background are initialized with the annotation of video sequences.Then,color and depth segmentation map can be computed by Bayes rule.Lastly,after the target background segmentation map is fused by the cosine window,this map is mask on the extracted features for reweighting,the color and depth histograms are updated by a linear interpolation after processing current frames.Compared with using only the pre-defined cosine window,the segmentation map can make the tracking algorithm pay more attention to the area belonging to the target,and the segmentation map can be dynamically adjusted during the tracking process.It is verified through simulation that the improved method is more robust than the tracking method using only the cosine window.In the Princeton RGBD tracking Benchmark,compared with the baselines,the accuracy of the proposed method is improved by 15.74%,and the success rate is increased by 13.52%.Secondly,a tracking algorithm that combines depth information in discriminative correlation filter framework is introduced.First,the K-means clustering method is used to process the one-dimensional depth histogram of bounding box in the depth image to obtain the spatial reliability map about the target shape.Then,the constrained correlation filter is computed according to spatial reliability map to avoid the boundary effects of conventional correlation filter.Lastly,the object is localized by summing per-channel responses weighted by the channel reliability scores in tracking stage.Occlusion is detected according to depth distribution of target region and responses of correlation filters.The model is not updated during occlusion,reducing drift problems.Scale is estimated by depth information of target.Simulation were performed on the Princeton RGBD tracking Benchmark and STC Benchmark.In the Princeton RGBD tracking Benchmark,compared with the CSR-DCF algorithm,the accuracy of the proposed method is improved by 25.25%,and the success rate is increased by 20.77%.The results show that the proposed method can track the target better under occlusion and scale changes scene.