Research On Visual Tracking Method For Complicated Environment

As an important part of the smart city,intelligent monitoring system is becoming more and more widely used in daily life.In the intelligent monitoring system,the most import thing is how to effectively apply intelligent video analysis technologies,and the visual tracking technology is a most basic one.Therefore,the research of visual tracking is essential to development of intelligent video analysis technology and the smart city in human society.In recent years,thanks to the development of various technologies including machine learning and deep learning,visual tracking has received much progress,and many algorithms with excellent performance have emerged.In general,these algorithms can be roughly grouped into two categories,traditional machine learning based ones and deep learning based ones.These kinds of methods have their own advantages.The traditional machine learning based visual tracking methods does not require a large number of samples to train a classifier,and usually achieve a considerable speed.But the major drawback is that the hand-craft features are not robust enough to some challenge scenarios.The major advantage of visual tracking methods using deep learning is that the network can extract more powerful features of target,but it needs a lot of samples for model training.However,most of these visual tracking algorithms are based towards how to design a robust appearance model and build effective tracking algorithms.However,various challenges in the process of visual tracking often have direct impacts on the final tracking performance,such as background clutter and partial occlusion.To handle these problems,we propose two novel approaches to visual tracking,and the major works and contributions are as follows:(1)We propose a spatially regularized graph learning algorithm to learn a robust object representation with suppression of background clutter for visual tracking.Specifically,we first divide the bounding box of the target object into a set of non-overlapping image patches,and assign each image patch with a weight to describe its importance belonging to the target object.Taking image patches as nodes,we learn a graph au.tomatically by using both global and local relationship among patches.The optimized weights of the graph nodes and the feature of the image patches are then combined to form the final target feature representation to reduce the influence of the background clutter.Finally,the target feature representation is used in the structured SVM to carry out object tracking.We have conducted extensive experiments on large-scale benchmark datasets to validate the effectiveness and robustness of the proposed approach.(2)To address the imaging limitations in single kind of sensors,we propose a general approach to fuse visible and thermal infrared information for robust visual tracking.Our approach relies on a novel two-stream convolutional neural network(CNN).First,we employ a basic CNN to extract powerful features from RGB and thermal sources to represent generic RGB and thermal representations of target object.For adaptive fusion of different modalities while avoiding redundant noises,the FusionNet is proposed to select most discriminative feature maps from the outputs of the two-stream CNN.Finally,we employ the multi-channel correlation filter to achieve robust and high-speed visual tracking.The effectiveness,robustness and efficiency of the proposed approach are verified by comparing with the existing methods on the benchmark dataset.