基于粒子濾波的自適應(yīng)目標跟蹤算法研究
發(fā)布時間：2018-11-09 14:05
【摘要】：隨著社會的信息化水平日益提高,傳統(tǒng)產(chǎn)業(yè)開始利用信息技術(shù)來提高生產(chǎn)效率、減少人力消耗,而計算機視覺技術(shù)已經(jīng)越來越多的被應(yīng)用于民用領(lǐng)域和軍事領(lǐng)域,生物特征識別、智能監(jiān)控、無人駕駛、智能武器等新興的概念開始不斷升溫。其中,視頻目標跟蹤技術(shù)是計算機視覺領(lǐng)域中的一個經(jīng)典研究課題,但是由于實際場景中往往存在光照變化、運動狀態(tài)突變、目標遮擋、相似物體干擾等復(fù)雜情況,當前已有的目標跟蹤技術(shù)仍難以滿足實際應(yīng)用的需求。目標跟蹤問題可以看作是由感興趣目標先前得知的位置來預(yù)測其在后續(xù)視頻序列中的空間位置,這是一個根據(jù)先驗條件來對當前狀態(tài)進行估計、驗證的過程,因此可以利用貝葉斯狀態(tài)估計的思想來對問題進行求解。本文正是對于其中最為經(jīng)典的粒子濾波算法進行研究,探討了視頻目標跟蹤中的一些關(guān)鍵性問題,主要的創(chuàng)新工作與研究成果包括以下幾方面:1.針對傳統(tǒng)粒子濾波目標跟蹤方法中粒子的多樣性不足以及易受場景干擾的問題,提出一種改進的免疫粒子濾波目標跟蹤方法,該方法基于人工免疫算法的思想,根據(jù)目標跟蹤中的關(guān)鍵性問題加入了抗體記憶庫、粒子集可信度判定等過程,以提高算法在較復(fù)雜場景中的魯棒性。2.建立合理的目標模型是粒子集更新結(jié)果趨向于目標狀態(tài)真實值的重要前提,本文針對傳統(tǒng)算法中的單一目標模型適應(yīng)性較差的問題,提出了加入自適應(yīng)學習機制的外觀模型與運動模型,同時利用了特征分片、背景權(quán)重等思想,并且給出了相應(yīng)的似然性計算方法。3.針對單目標粒子濾波跟蹤方法直接應(yīng)用到多目標跟蹤問題時易出現(xiàn)的問題,提出了一個快速的交互目標判定與匹配算法,該方法適用于粒子濾波框架下的跟蹤方法,可以在一定程度上提高多目標跟蹤的準確性。本文嘗試通過對傳統(tǒng)的粒子濾波目標跟蹤算法進行改進,使其在較為復(fù)雜的實際場景中提高性能。分別在Visual Tracker Benchmark測試庫、PETS 2009Benchmark Data測試庫以及車載相機拍攝的動態(tài)場景中選擇了多段典型的視頻進行算法的對比實驗與分析,通過L1-偏差、目標區(qū)域覆蓋比、多目標跟蹤精確度、算法運行速度等統(tǒng)計指標驗證了所提算法較傳統(tǒng)方面具有明顯的提高,在實際場景中達到了較好的適應(yīng)性、魯棒性和實時性。
[Abstract]:With the increasing level of information technology in society, traditional industries begin to use information technology to improve production efficiency and reduce human consumption, and computer vision technology has been more and more used in civilian and military fields. New concepts such as biometric identification, intelligent surveillance, driverless and intelligent weapons are starting to heat up. Among them, video target tracking technology is a classical research topic in the field of computer vision. However, because of the complex situation, such as illumination change, moving state mutation, object occlusion, similar object interference and so on, in the actual scene, video target tracking technology often exists in the field of computer vision. The existing target tracking technology is still difficult to meet the needs of practical applications. The target tracking problem can be regarded as predicting the spatial position of the object of interest in the subsequent video sequence from the position previously known, which is a process of estimating and verifying the current state according to a priori condition. Therefore, Bayesian state estimation can be used to solve the problem. In this paper, the most classical particle filter algorithm is studied, and some key problems in video target tracking are discussed. The main innovative work and research results include the following aspects: 1. Aiming at the shortage of particle diversity and the vulnerability to scene interference in traditional particle filter target tracking methods, an improved immune particle filter target tracking method is proposed, which is based on the idea of artificial immune algorithm. According to the key problems in target tracking, the antibody memory library and particle set reliability evaluation are added to improve the robustness of the algorithm in more complex scenarios. 2. Establishing a reasonable target model is an important prerequisite for updating the result of particle set towards the real value of target state. This paper aims at the problem of poor adaptability of single objective model in traditional algorithm. The appearance model and motion model with adaptive learning mechanism are put forward, and the idea of feature segmentation and background weight are used, and the corresponding likelihood calculation method is given. 3. Aiming at the problem that single target particle filter tracking method is easy to appear when it is directly applied to multi-target tracking problem, a fast interactive target determination and matching algorithm is proposed, which is suitable for tracking method under particle filter framework. The accuracy of multi-target tracking can be improved to some extent. This paper attempts to improve the traditional particle filter target tracking algorithm to improve its performance in more complex practical scenarios. In the dynamic scene of Visual Tracker Benchmark test library, PETS 2009Benchmark Data test library and vehicle camera, we choose several typical video to carry on the contrast experiment and analysis, through L1-deviation, coverage ratio of target area, precision of multi-target tracking. The algorithm running speed and other statistical indicators verify that the proposed algorithm has obvious improvement compared with the traditional algorithm, and achieves better adaptability, robustness and real-time performance in the actual scene.
【學位授予單位】：吉林大學
【學位級別】：碩士
【學位授予年份】：2017
【分類號】：TP391.41
[Abstract]:With the increasing level of information technology in society, traditional industries begin to use information technology to improve production efficiency and reduce human consumption, and computer vision technology has been more and more used in civilian and military fields. New concepts such as biometric identification, intelligent surveillance, driverless and intelligent weapons are starting to heat up. Among them, video target tracking technology is a classical research topic in the field of computer vision. However, because of the complex situation, such as illumination change, moving state mutation, object occlusion, similar object interference and so on, in the actual scene, video target tracking technology often exists in the field of computer vision. The existing target tracking technology is still difficult to meet the needs of practical applications. The target tracking problem can be regarded as predicting the spatial position of the object of interest in the subsequent video sequence from the position previously known, which is a process of estimating and verifying the current state according to a priori condition. Therefore, Bayesian state estimation can be used to solve the problem. In this paper, the most classical particle filter algorithm is studied, and some key problems in video target tracking are discussed. The main innovative work and research results include the following aspects: 1. Aiming at the shortage of particle diversity and the vulnerability to scene interference in traditional particle filter target tracking methods, an improved immune particle filter target tracking method is proposed, which is based on the idea of artificial immune algorithm. According to the key problems in target tracking, the antibody memory library and particle set reliability evaluation are added to improve the robustness of the algorithm in more complex scenarios. 2. Establishing a reasonable target model is an important prerequisite for updating the result of particle set towards the real value of target state. This paper aims at the problem of poor adaptability of single objective model in traditional algorithm. The appearance model and motion model with adaptive learning mechanism are put forward, and the idea of feature segmentation and background weight are used, and the corresponding likelihood calculation method is given. 3. Aiming at the problem that single target particle filter tracking method is easy to appear when it is directly applied to multi-target tracking problem, a fast interactive target determination and matching algorithm is proposed, which is suitable for tracking method under particle filter framework. The accuracy of multi-target tracking can be improved to some extent. This paper attempts to improve the traditional particle filter target tracking algorithm to improve its performance in more complex practical scenarios. In the dynamic scene of Visual Tracker Benchmark test library, PETS 2009Benchmark Data test library and vehicle camera, we choose several typical video to carry on the contrast experiment and analysis, through L1-deviation, coverage ratio of target area, precision of multi-target tracking. The algorithm running speed and other statistical indicators verify that the proposed algorithm has obvious improvement compared with the traditional algorithm, and achieves better adaptability, robustness and real-time performance in the actual scene.
【學位授予單位】：吉林大學
【學位級別】：碩士
【學位授予年份】：2017
【分類號】：TP391.41
【參考文獻】
相關(guān)期刊論文 前9條
1 李文輝;陳昱昊;王瑩;;自適應(yīng)的免疫粒子濾波車輛跟蹤算法[J];吉林大學學報(理學版);2016年05期
2 林濤;劉以安;;自適應(yīng)蟻群算法在多目標跟蹤中的應(yīng)用[J];計算機仿真;2014年09期
3 肖沁雨;;智能視頻監(jiān)控關(guān)鍵技術(shù)分析[J];制造業(yè)自動化;2012年12期
4 蔣戀華;甘朝暉;蔣e，
本文編號：2320631
本文鏈接：http://www.sikaile.net/kejilunwen/ruanjiangongchenglunwen/2320631.html
最近更新
教材專著
