Research On Tracking Methods Of Moving Target In Complicated Conditions

Moving target tracking is one of the most important contents in computer vision, it has a wide range of applications, such as intelligent surveillance, precision guide and visual navigation of robots. Although moving target tracking has been studied by researchers at the domestic and abroad, it remains several challenging problem in complicated conditions,especially for pose variation, illumination changes, scale variation, background clutters, as well as occlusion. In the dissertation, the research is focused on target tracking algorithm based on machine learning, the main works are summarized as follows:(1) An efficient Tracking-Learning-Detection(TLD) target tracking algorithm based on BInary Normed Gradients(BING) is proposed. The local tracker failure predicting method based on spatio-temporal context is introduced as well as the estimation algorithm for the global motion model to improve the precision of the tracker. Replacing a sliding window for search the target, BING algorithm is combined with the cascaded classifier to detect the candidate target, thus reducing the search space and improving the processing speed of the detector under the premise of ensuring detection precision. The sample weight is integrated into the online learning procedure to improve the accuracy of the classifier and alleviate drift to some extent. The experimental results show that compared with the state-of-the-art tracking algorithm, the proposed algorithm has the superior performance on tracking precision and tracking speed in complicated conditions.(2) A novel algorithm for target tracking based on deep learning is proposed under TLD framework. The target is predicted and tracked by the enhanced flock of tracker. Every sample is weighted by P-N learning to improve the precision of the classifier. A deep detector is composed of the stacked denoising autoencoder and sigmoid classifier, combines with multi-scale scanning window with global search strategy to detect possible targets. The unsupervised feature learning is used by a stacked denoising autoencoder to optimize the network parameters, and transfers the parameter learned to the online tracking task to extract the features. This alleviates the problem of not having much labeled data in visual tracking applications. A classification neural network is used to distinguish the tracked object from the background on the online tracking process. The further tuning is allowed to adapt to the appearance changes of the moving object through the supervised learning. To reduce the matching template and the computational complexity, the online template-based object model is clustered in binary search tree using k-means. The experimental results show that compared with the state-of-the-art tracking algorithm, the proposed algorithm has moreaccuracy and better robust in complicated conditions.(3) A novel algorithm for target tracking based on incremental deep learning is proposed. The particle set is distributed by particle filter. To complete the image representations expressively, the features of particle area are extracted by a stacked denoising autoencoder which is obtained by the unsupervised feature learning. The incremental feature learning optimizes the feature size to adapt the appearance changes of the target and obtain the compact feature representation. The learning algorithm is consisted of adding features and merging features. It adds new features to introduce the new information, and merges similar features to obtain compact feature representation. A linear support vector machine classifies the optimized feature set to get the confidence of the particle and fine tune the deep network. The tracking result is the particle with the largest confidence. To solve the problem of particle degradation and depletion, the double-resampling method is introduced which adapts the particle size. The experimental results show that compared with the state-of-the-art tracking algorithm, the proposed algorithm has the superior performance on tracking precision and robustness in complicated conditions.(4) For the problem of the target tracking algorithm based on Multiple Instance Learning(MIL), a novel algorithm for target tracking based on multiple instance deep learning is proposed. In original MIL algorithm, the Haar-like feature can not represent the image effectively and is easy to fail when influenced by the outside environment. A stacked denoising autoencoder is used to learn image features and express the image representations obtained effectively to adapt to the dynamic change of the environment. Selected feature vector could not be replaced in the original MIL algorithm, which has difficulty reflecting the changes of the target and the background. Thus, some weakest discriminative feature vector is replaced when weak classifiers are selected to adapt to the dynamic changes of the target. Aiming at the deficiency of using motion model where the location of the target is likely to appear within a radius in original MIL algorithm, the particle filter estimates object’s location to increase the tracking precision. The experimental results show that compared with the original MIL algorithm and other state-of-the-art tracking algorithm, the proposed algorithm improves the tracking accuracy and the robustness in complicated conditions.