Robust Feature Learning For Object Tracking

At present,the construction of smart city is in full swing,and intelligent monitoring,as an important part of it,is urging the continuous progress of computer vision.Object tracking is the cornerstone of computer vision,which plays an important role in improving the reliability of the intelligent monitoring system and helps to construct a safe society.In recent years,object tracking has already had many breakthroughs based on significant progress of machine learning and deep learning.However,there are still some challenges need to be solved.On the one hand,learning a robust object representation is very critical for improving tracking accuracy.On the other hand,object tracking requires real-time tracking speed in most applications,and thus the tracking speed is also a core part of tracking performance evaluation.To deal with these problems,this dissertation proposes two tracking approaches to learn a robust object representation efficiently.The main works are as follows.Since most of current object tracking methods use a bounding box to locate the target,it inevitably introduces background information in describing the target,which may cause model drift.Therefore,a general method to learn a robust object representation for tracking is proposed in this dissertation,which relies on a novel patch-based absorbing Markov chain algorithm.Specifically,the bounding box is first divided into a plurality of non-overlapping image patches,and these image patches are taken as nodes.Then a sparse-to-full manner is employed to learn a dense affinity matrix and the matrix will be applied into the absorbing Markov chain.The absorbed time between transient nodes and absorbing nodes in an absorbing Markov chain represents the similarities between these two kinds of nodes,and thus the foreground and background probability distribution of image patches can be obtained by calculating the absorbed time after setting foreground and background seeds as absorbing nodes respectively based on a priori hypothesis.Next,these two kinds of probability distributions are further used to calculate the foreground weights of the image patches and the weights are integrated into the patch-based feature description,which can reduce the interference of background noise and achieve a robust object representation.Finally,the weighted feature description is input into the structured support vector machine algorithm for tracking.In addition,this dissertation also designs a strategy of foreground seeds optimization to remove unreliable foreground seeds from the prior hypothesis so as to improve the accuracy of weight calculation.Experimental results on the public benchmark dataset demonstrate the effectiveness of our method.Deep learning tracking methods generally use only high-level semantic features to locate target objects and ignore detailed information from low-level deep features.Therefore,the problems of background clutter and inaccurate positioning may occur.This dissertation proposes a multi-level feature aggregation Siamese network,which effectively integrates the complementary deep features of different layers.The proposed method can improve the tracking accuracy while maintaining a real-time tracking speed.First,a lightweight multi-level feature aggregation module is introduced during feature extraction.It aggregates feature maps from multiple scales into a unified space at the same resolution and combines both local and global context information without increasing too much computational burden.Second,the architecture of Siamese network is used to learn the similarity between the template region and the candidate region.The candidate region which is most similar to the template is predicted to be the target position.Experimental results on the public benchmark dataset demonstrate that multi-level feature aggregation achieves a robust object representation,and enables us to construct a Siamese tracker with a balanced tracking accuracy and tracking speed in an end-toend training manner.