Font Size: a A A

Research On Visual Tracking Algorithms Based On Feature Covariance And Particle Filter

Posted on:2016-12-05Degree:DoctorType:Dissertation
Country:ChinaCandidate:Z H FanFull Text:PDF
GTID:1108330488457662Subject:Pattern Recognition and Intelligent Systems
Abstract/Summary:PDF Full Text Request
With the increasing improvement of computer, image sensor, memory device, etc., the technique of visual tracking plays a more and more important role in the practical applications, such as the electro-optical weapon guidance system, the security surveillance system, etc., becoming an important and difficult problem in the research fields of computer vision and artificial intelligence. The task of visual tracking is to sequentially estimate the positions and the scales of the targets in the image sequence based on the priori information about the apparent features and the motion model of the target. In order to achieve a long-time, stable, accurate and efficient tracking in a complex environment, we need to make efforts in the following aspects:(a) For the apparent feature description, we should select the features according to the characteristics of the image data, and thus establish an effective multi-feature description model.(b) In the design of the tracking algorithm, we need to weaken the clutter and improve the estimation accuracy.(c) For real-world applications, we should ensure the run time of the entire algorithm within an acceptable range. For these purposes, the dissertation utilizes feature covariance matrices and particle filter(PF) to carry out the research on visual tracking methods, and focuses on solving the key problems of feature descriptions, intelligent scenario analysis and filtering algorithms. The main contributions of the dissertation are as follows:1. The visible color target tracking encounters the problem of how to establish an effective multi-feature description model. For this reason, a visual target tracking algorithm based on the HSV(Hue saturation value) color feature covariance matrices is proposed. In the proposed algorithm, the HSV color features are fused into the feature covariance matrices which are combined with PF to accomplish the robust and accurate tracking. Experimental results show that the HSV color feature covariance matrices can make full use of the color information of the target, and have the advantages of high information independence and low dimensions, obtaining an effective description of the visible color target.2. The infrared target tracking is faced with the problems of how to utilize the significance of the infrared target and resist the interference. To handle these, an infrared target tracking algorithm using the scenario analysis of the activated region is proposed. The proposed algorithm segments the target from the background by using the significance of the high brightness of the infrared target, then obtains the position measurements, the scale measurements and the interference status based on the scenario analysis of the activated region. And the auxiliary information can improve the tracking performance. Experimental results demonstrate that the scenario analysis of the activated region improves the reliability of the weights of the particles, the adjustment accuracy of the tracking window and the updating strategy of the feature template, forming a stable and reliable mechanism.3. Particle filter is confronted with the problems of the sampling efficiency and interference suppression. For this purpose, a novel iterative particle filter(IPF) tracking algorithm is proposed. By sampling the particles iteratively with the search scope annealed, the proposed IPF can converge to the true target state as close as possible. Experimental results show that IPF can improve the sampling efficiency and weaken the impact of the clutter on the fusion estimation, obtaining the effective, accurate and robust tracking performance.4. Particle filter is also faced with the problem of the spatial coverage of particles, a novel box particle filter visual tracking(BPF-VT) algorithm is proposed. Considering that the measurement function is a non-elementary function and the coordinates of the digital image are discrete in visual tracking, the proposed BPF-VT improves the measurement interval mapping, the contraction method and the resampling strategy to make the box particles effectively cover the state space. Experimental results demonstrate that BPF-VT can improve the sampling efficiency and the computational speed. Especially, the proposed BPF-VT has good tracking performance for the moving target with large dynamic range.
Keywords/Search Tags:Visual tracking, Particle filter(PF), Feature covariance matrices, HSV, Scenario analysis, Annealing, Box particle, Sampling efficiency
PDF Full Text Request
Related items