Integrating Multiple Visual Cues For Researching On Single Target Tracking

Visual object tracking is a popular but difficult problem in the field of computer vision.It has garnered significant attention due to its widespread use in applications such as surveillance,driverless cars and medical diagnostic systems.However,practical tracking applications are faced with various complex environment scenarios and visual object tracking remains a challenging problem.Therefore,it is of great significance to explore efficient and robust tracking algorithms.To this end,we propose the corresponding solutions to improve the tracking performance.One is based on correlation filter tracking and another is built upon siamese network.Extensive experiments investigating online visual object tracking benchmarks show the effectiveness of the proposed method.The specific research work includes the following two aspects:(1)Correlation filter-based trackers have shown favorable tracking performance in online benchmark datasets.The current methods often assume that the target shape is well approximated by an axially aligned rectangle which is a single-center gaussian distribution.However,most of the targets being tracked are not strictly rectangular shaped and the bounding box representation often inevitably incorporates multiple pieces of background information into the model,this may lead to model drift and tracking failure.To address this problem,we propose a method of correlation filter tracking by integrating multiple background cues to train two separate filters.Firstly,the first filter learns from all training examples and generates the target response point.Then other response points are obtained by means of gaussian interpolation.All the response points generated by the first filter are regarded as the prior target responses and used to train the second filter.While the second filter makes use of the results from the first filter so as to update the regression to a more reliable target response.Additionally,we incorporate a novel model update strategy into our formulation to further improved the performance of the algorithm.Extensive experiments investigating online visual object tracking benchmarks show the effectiveness of the proposed method.(2)Some trackers based on siamese network often fail to make full use of the complementary properties between deep features and shallow features.Meanwhile,background information is introduced in the process of feature extraction,which affects the tracking performance.To address this problem,we exploit multi-cue cascades for building a robust end-to-end visual tracking.Firstly,the pre-trained convolutional neural network is used to extract the depth features and combine with the shallow features to learn the response of each level respectively.Then it is embedded into the dynamic siamese networks for end-to-end training.In order to reduce the influence of background interference,regularization constraints are added in the learning process of high-level features to further suppress the influence of background information.Finally,the feature response of each layer is fused to obtain the final tracking result.Extensive experiments demonstrate that the proposed tracker performs well,while being more robust for background clutters.