Visual Object Tracking Based On Occlusion Detection And Deep Siamese Network

Visual object tracking technology has a wide range of applications and is therefore an important research topic in the field of computer vision.The outline of the target tracking algorithm is as follows.In the first frame of the video,the size and position of the target is given by the bounding box surrounding the target,which is used to initialize the target template.In subsequent frames,the tracking algorithm first locates a group of target candidates according to the possible positions of the target,and extracts features of these candidates.The algorithm estimates the probability of candidates being the target.The candidate with the greatest probability is treated as the target in current frame.Finally,the template update module updates the target template according to the target information of current frame.Target template updates are necessary because the appearance of the target may vary during the tracking process.However,when the target is occluded by the background,it will lead to the wrong update of the target template if the tracking algorithm cannot accurately detect occlusion.This may reduce the capability of the target template to discriminate the target and the background and causing the algorithm to lose the target in subsequent frames.In order to enhance the robustness of the target tracking algorithms to occlusion,we propose a Context-based Occlusion Detection method(COD).The method divides the background around the target in the current frame to patches,and tracks each background patch separately.Then it detects the background patches that overlap with the target with high tracking reliability.These patches are considered as occluders,whose number indicates the severity of the occlusion.The COD method can effectively detect the occurrence of occlusion.This method can be combined with any existing target tracking algorithm.Therefore,we proposes a tracking framework where the COD method is used as a sub-module for occlusion detection.When the COD method detects heavy occlusion,the target template will stop updating.This paper implements a correlation filter based tracking algorithm DSST and combines it with the COD method.Experiments on benchmarks such as OTB-2013 and VOT-2015 show that the COD method can significantly improve the performance of the target tracking algorithm.We further improves COD approach and proposes a Context-based Adaptive Occlusion Detection method(CAOD).The COD method does not have the ability to adapt to the target's scale.Unlike COD,in CAOD,the number of background patches will be determined by the target's scale.Meanwhile,in CAOD the proportion of the area being occluded is used to estimate the degree of occlusion.In order to evaluate the performance of various tracking approaches under occlusion,we establish an occlusion video dataset.The videos of this dataset are specially processed to evaluate the tracking accuracy and robustness of the tracking algorithm in the case of occlusion.A number of comparison experiments are carried out on this dataset,which fully confirms that the CAOD algorithm can improve the robustness of the tracking algorithm against occlusion.Deep learning greatly improves the accuracy of the target tracking algorithm,and the tracking algorithm based on the Siamese network structure is a proper method.This paper implements a tracking method based on deep Siamese network structure.After feature extraction is performed on the possible region of the target,the cross-correlation operation is applied on the target template and each candidate to obtain the response map.The peak location is selected as the position of the target.This method avoids the training on the large-scale tracking dataset,and only needs the pre-trained convolutional neural network.The structure is simple but the performance is good.The main contributions of this article can be summarized as follows:1.A context-based occlusion detection method(COD)is proposed,and a target tracking framework based on occlusion detection is developed,which makes the target tracking algorithm robust against occlusion.2.A context-based adaptive occlusion detection method(CAOD)is proposed.An occlusion video benchmark is built,which is specifically used to evaluate the performance of the tracking algorithm under occlusion.3.A target tracking algorithm based on Siamese network structure is proposed.The algorithm has simple structure,fast running speed and good tracking accuracy.