本文選題：目標跟蹤　切入點：機器學習　出處：《寧夏大學》2017年碩士論文

【摘要】：目標跟蹤技術(shù)是計算機視覺領(lǐng)域重要技術(shù)之一,在不同領(lǐng)域都起到一定作用。目標跟蹤技術(shù)的不斷進步,也在不斷提升人們的生活質(zhì)量。一方面,傳統(tǒng)的目標跟蹤算法,在指定的場景、目標下往往能起到比較顯著的效果,不論是時效性還是準確率,都有一定的優(yōu)勢,從長時間跟蹤的層面看,傳統(tǒng)目標跟蹤算法的誤差不斷累積。另一方面,傳統(tǒng)的目標跟蹤算法在場景和目標上具有局限性,無法應用到其他目標或場景中。隨著計算機的不斷智能化,大數(shù)據(jù)時代的到來,無疑給機器學習帶來了新的發(fā)展機遇。機器學習屬于多領(lǐng)域交叉學科,從大量數(shù)據(jù)中學習產(chǎn)生新知識,并對未知事物進行預測。機器學習理論在目標跟蹤上的應用,通過將目標跟蹤問題轉(zhuǎn)化為數(shù)據(jù)分類問題,簡化了目標跟蹤的復雜性,適用于各類復雜的目標跟蹤問題。本文對目標跟蹤算法進行研究,主要研究TLD算法,發(fā)現(xiàn)其在長時間單一目標跟蹤問題上具有較高的準確率,但在運行速度上無法達到實時跟蹤的效果。本文在TLD算法的框架基礎(chǔ)上,提出了基于幀間差分法的TLD跟蹤算法,經(jīng)過大量實驗數(shù)據(jù)表明,其在跟蹤速度上較原算法提高了 3.9倍,因而本文將該算法應用到目標跟蹤系統(tǒng),并通過目標檢測模塊實現(xiàn)非監(jiān)督的基于機器學習的目標跟蹤系統(tǒng)。本文開展的具體研究工作如下:1、介紹了目標跟蹤的發(fā)展歷史、研究現(xiàn)狀。介紹了幀間差分法、STC算法和TLD算法的原理,從理論上分析三者的優(yōu)劣勢;2、對比傳統(tǒng)目標跟蹤算法與基于機器學習的目標跟蹤算法,仿真實現(xiàn)幀間差分法、STC算法、TLD算法在目標跟蹤上的應用,并對結(jié)果進行對比分析;3、對TLD算法進行深入研究,分析各模塊的運行機制,提出了幀間差分法的TLD跟蹤算法,應用到目標跟蹤系統(tǒng),并進行大量實驗,展示系統(tǒng)在視頻圖像跟蹤的準確性;4、設(shè)計并完成了基于機器學習的目標跟蹤系統(tǒng),將本文所提出算法應用到目標跟蹤系統(tǒng)中,實現(xiàn)其應用價值,并將系統(tǒng)移植到Linux系統(tǒng),最后在此平臺上進行實驗,驗證其跟蹤效果。
[Abstract]:Target tracking technology is one of the important technologies in computer vision field, which plays a certain role in different fields. With the development of target tracking technology, people's quality of life is also being improved. On the one hand, traditional target tracking algorithm, In the specified scene, the target can often play a more significant effect, whether it is timeliness or accuracy, have certain advantages, from the long-term tracking level, the traditional target tracking algorithm error accumulation, on the other hand, The traditional target tracking algorithm is limited in scene and target, and can not be applied to other target or scene. There is no doubt that machine learning brings new opportunities for development. Machine learning belongs to a multi-domain interdisciplinary subject, learning from a large number of data to produce new knowledge, and to predict the unknown. The application of machine learning theory in target tracking. By transforming the target tracking problem into a data classification problem, it simplifies the complexity of target tracking and is suitable for all kinds of complex target tracking problems. In this paper, the target tracking algorithm is studied, and the TLD algorithm is mainly studied. It is found that it has high accuracy in long time single target tracking problem, but it can not achieve the effect of real-time tracking in running speed. In this paper, based on the frame of TLD algorithm, a TLD tracking algorithm based on inter-frame difference method is proposed. A large number of experimental data show that the tracking speed of the algorithm is 3.9 times higher than that of the original algorithm, so the algorithm is applied to the target tracking system in this paper. An unsupervised target tracking system based on machine learning is implemented through target detection module. The specific research work in this paper is as follows: 1. The development history of target tracking is introduced. This paper introduces the principle of inter-frame difference algorithm and TLD algorithm, analyzes their merits and demerits in theory, and compares the traditional target tracking algorithm with that based on machine learning. The application of TLD algorithm in target tracking is realized by simulation, and the results are compared and analyzed. The TLD algorithm is studied deeply, and the running mechanism of each module is analyzed, and the TLD tracking algorithm of inter-frame difference method is proposed. It is applied to the target tracking system, and a large number of experiments are carried out to show the accuracy of the system in video image tracking. A target tracking system based on machine learning is designed and completed. The algorithm proposed in this paper is applied to the target tracking system. The application value is realized, and the system is transplanted to Linux system. Finally, experiments are carried out on this platform to verify its tracking effect.
【學位授予單位】：寧夏大學
【學位級別】：碩士
【學位授予年份】：2017
【分類號】：TP391.41;TP181

【參考文獻】

相關(guān)期刊論文 前10條

1 管皓;薛向陽;安志勇;;在線單目標視頻跟蹤算法綜述[J];小型微型計算機系統(tǒng);2017年01期

2 楊豐瑞;杜奎;莊園;;TLD目標跟蹤算法綜述[J];電視技術(shù);2016年10期

3 尹宏鵬;陳波;柴毅;劉兆棟;;基于視覺的目標檢測與跟蹤綜述[J];自動化學報;2016年10期

4 焦蓬斐;秦品樂;苗啟廣;劉毛毛;呂國宏;;基于多新息Kalman濾波的TLD改進算法[J];數(shù)據(jù)采集與處理;2016年03期

5 劉威;趙文杰;李成;;時空上下文學習長時目標跟蹤[J];光學學報;2016年01期

6 邢藏菊;溫蘭蘭;何蘇勤;;TLD視頻目標跟蹤器快速匹配的研究[J];小型微型計算機系統(tǒng);2015年05期

7 呂枘蓬;蔡肖芋;董亮;涂繼輝;;基于TLD框架的上下文目標跟蹤算法[J];電視技術(shù);2015年09期

8 高仕博;程詠梅;肖利平;韋海萍;;面向目標檢測的稀疏表示方法研究進展[J];電子學報;2015年02期

9 黃凱奇;陳曉棠;康運鋒;譚鐵牛;;智能視頻監(jiān)控技術(shù)綜述[J];計算機學報;2015年06期

10 屈晶晶;辛云宏;;連續(xù)幀間差分與背景差分相融合的運動目標檢測方法[J];光子學報;2014年07期

，

本文編號：1669625