基于檢測的數據關聯(lián)多目標跟蹤算法研究
發(fā)布時間：2019-06-05 15:15
【摘要】：視頻序列的多目標跟蹤研究是計算機視覺領域的一個重要內容,已經廣泛應用于國防、視頻監(jiān)控、智能導航/輔助駕駛、智能機器人、行為分析、視頻檢索、生物醫(yī)學等領域。視頻多目標跟蹤的目的是在視頻序列中標定出各個目標的運動軌跡。然而,受成像質量下降、噪聲和背景干擾、目標外觀和運動模式的變化、被跟蹤目標數目的不確定性、復雜多變的遮擋等諸多因素的影響,多目標跟蹤算法研究是一個頗具挑戰(zhàn)性的課題,還面臨大量的理論和技術問題有待解決。近十年來,隨著目標檢測器性能的不斷提升,基于檢測的目標跟蹤方法引起了廣泛的關注,已成為當前主流的多目標跟蹤方法。這類方法將檢測器輸出的檢測響應作為輸入,通過數據關聯(lián)技術將屬于同一目標的檢測響應逐一連接、最終獲得各個目標的運動軌跡。關聯(lián)模型設計是基于檢測的數據關聯(lián)技術的關鍵,一個好的關聯(lián)模型應該盡可能的融合那些反映目標軌跡內在屬性的觀測信息,在噪聲和復雜場景下提供可靠的軌跡關聯(lián)。本文圍繞多目標跟蹤領域前沿發(fā)展動態(tài),著重對關聯(lián)模型及其在多目標跟蹤方法中的應用進行了深入研究。主要工作包括以下四部分：(1)基于霍夫森林學習的多目標跟蹤方法多級關聯(lián)是一種有效的多目標跟蹤方法,本文在逐級關聯(lián)框架下,提出了一種基于霍夫森林分類器的多目標跟蹤算法。首先通過保守關聯(lián)算法生成可靠的短小軌跡片；隨后以逐級處理方式從跡片集上提取具有判別性能的外觀和運動特征,生成訓練樣本并構建霍夫森林；在測試階段,利用森林葉子節(jié)點中存儲的有效碼元信息去估計軌跡片間的連接概率,最終將軌跡關聯(lián)轉化為最大后驗概率準則(Maximum-A-Posterior, MAP)下的求解問題。實驗證明了基于霍夫森林的軌跡片關聯(lián)模型的有效性：與一些國外同行的近期算法相比較,該方法取得了與之相當的跟蹤效果(2)基于遮擋推理模型的孤立響應點匹配由于運動場景的復雜性、頻繁發(fā)生的遮擋等,即便是目前最先進的檢測器也存虛檢、漏檢、檢測不精確等問題；可靠軌跡片生成階段所采取的保守關聯(lián)策略也會遺漏一些檢測響應。上述問題都將導致在最后的跟蹤結果中,出現(xiàn)不能和任何軌跡相關聯(lián)的孤立響應點,從而使得目標軌跡間隙增大、平滑性下降。針對此問題,本文提出一種新的遮擋推理模型,以此推斷出遮擋目標的被遮擋區(qū)域和非遮擋區(qū)域；在此基礎上設計被遮擋目標的融合特征描述,提出孤立響應點與目標軌跡間的匹配策略,有效解決了孤立響應點的目標歸屬問題。作為一種填補軌跡間隙的后處理技術,本章方法對于采用軌跡片關聯(lián)的跟蹤算法具有普適性。(3)基于霍夫森林條件隨機場的多目標跟蹤方法基于條件隨機場(conditional random field, CRF)的多目標跟蹤近年來已成為一個研究熱點。作為CRF模型的核心,CRF模型參數估計和狀態(tài)推理非常棘手,傳統(tǒng)上通常采用近似算法或啟發(fā)式方法對參數進行估計,而CRF推理過程也易陷入局部最優(yōu)。本文提出霍夫森林條件隨機場模型(Hough Forest Conditional Random Field ,HFRF),該方法通過SW-cuts算法計算MH跳轉接受概率以實現(xiàn)狀態(tài)推理求解,而利用霍夫森林提供CRF推理所需要的概率參數；HFRF將CRF模型參數學習和推理嵌入到同一個框架中,從而規(guī)避了傳統(tǒng)CRF跟蹤方法中的難題。此外,與傳統(tǒng)的CRF圖模型不同,HFRF對每條邊額外定義了一個二元指示隱變量,將傳統(tǒng)CRF中的二元組結構關系擴展到三元,可以考慮更多運動目標的時空域關系,利用該三元組結構有助于跟蹤算法的優(yōu)化和性能提升(4)基于數據聯(lián)合表達的多目標跟蹤方法傳統(tǒng)的數據關聯(lián)模型大多依據數據間的差異性建模,例如計算兩個特征間的距離；這一操作本質上是降維過程,將導致原始數據可分性的部分丟失。針對此問題,本文在傳統(tǒng)CRF圖模型下,提出了一種基于數據聯(lián)合分布建模的多目標跟蹤算法。該方法構造二元勢函數表征軌跡片間的相關性,構造高階類別損失函數(正則項)用以約束求解的目標個數,在此基礎上得到代價方程,最終通過代價最小化實現(xiàn)CRF模型下的類別標定。其中軌跡片間的勢函數建模為兩個假設條件下的數據聯(lián)合分布,通過建立關聯(lián)數據的相容性、相斥性概率,完成CRF模型的推理過程。該方法另一個特點是利用霍夫森林葉子節(jié)點間存儲的樣本類別的分布特性,以無參的形式實現(xiàn)兩個假設條件下數據聯(lián)合分布概率的估計。在多個數據庫上進行的仿真實驗證明了本文方法的有效性；該方法所提出的關聯(lián)數據相容性、相斥性建模思想,為多目標跟蹤算法設計提供了一種新的途徑
[Abstract]:The multi-objective tracking of video sequences is an important content in the field of computer vision, and has been widely used in the fields of national defense, video surveillance, intelligent navigation/ assistant driving, intelligent robot, behavioral analysis, video retrieval, biomedicine and so on. The purpose of the video multi-target tracking is to calibrate the motion trajectory of each target in the video sequence. However, the study of multi-objective tracking algorithm is a challenging task, which is influenced by the image quality, noise and background interference, the target appearance and the motion pattern, the uncertainty of the number of tracked targets, and the complex and variable occlusion. There is also a large number of theoretical and technical problems to be solved. In the last ten years, with the continuous improvement of the performance of the target detector, the target tracking method based on the detection has attracted wide attention and has become the current multi-target tracking method. In the method, the detection response output by the detector is taken as an input, and the detection responses belonging to the same target are connected one by one by the data association technology, and finally the motion trail of each target is obtained. The design of the association model is the key to the data association technology based on the detection, and a good correlation model should be used to integrate the observation information reflecting the intrinsic attributes of the target track as much as possible, and provide a reliable track association under the noise and the complex scene. This paper focuses on the development of the multi-object tracking field, and focuses on the correlation model and its application in the multi-objective tracking method. The main work includes the following four parts: (1) The multi-objective tracking method based on the Hough forest learning is an effective multi-target tracking method. In this paper, a multi-target tracking algorithm based on the Hough forest classifier is proposed in this paper. the method comprises the following steps of: firstly, generating a reliable short track slice through a conservative correlation algorithm; then, extracting the appearance and the motion characteristic with the discrimination performance from the trace piece set in a step-by-step processing mode, generating a training sample and constructing a Hough forest; and in the testing phase, The connection probability between the track slices is estimated by using the effective symbol information stored in the forest leaf node, and finally the path association is converted into a solution problem under the maximum posterior probability criterion (MAP). The experiment proves the effectiveness of the correlation model of the track slice based on the Hough forest: compared with the recent algorithm of some foreign peers, the method has obtained the corresponding tracking effect (2) matching the complexity of the motion scene due to the isolated response point of the occlusion reasoning model, The frequently occurring occlusion, etc., even the most advanced detector at present, also has the problems of false detection, missed detection, inaccurate detection, etc. The conservative association strategy adopted in the reliable track slice generation phase will also miss some detection responses. The above problems will result in an isolated response point that cannot be associated with any trace in the final tracking result, so that the target track gap is increased and the smoothness decreases. In this paper, a new occlusion reasoning model is proposed, which can be used to infer the occluded area and the non-occlusion area of the occlusion target. Based on this, the fusion feature description of the occlusion target is designed, and the matching strategy between the isolated response point and the target track is proposed. And the problem of the target attribution of the isolated response point is effectively solved. As a post-processing technique to fill the gap of the track, the method of this chapter is of general applicability to the tracking algorithm associated with the track slice. (3) The multi-objective tracking method based on the conditional random field (CRF) has become a hot topic in recent years based on the multi-objective tracking method of the Hough forest condition with the airport. As the core of the CRF model, the parameter estimation and state-based reasoning of the CRF model is very difficult, and the parameters are estimated conventionally by using a heuristic algorithm or a heuristic method, and the CRF reasoning process is also easy to fall into a local optimal. In this paper, Hough Forest Conditional Random Field (HFRF) is proposed in this paper. The probability of MH jump acceptance is calculated by the SW-ctrl algorithm to realize the state-based reasoning, and the probability parameters required for CRF inference are provided by the Hough forest. HFRF has embedded CRF model parameter learning and reasoning in the same framework, thus avoiding the difficulty in the traditional CRF tracking method. In addition, unlike the traditional CRF map model, HFRF additionally defines a binary indicator hidden variable for each side, extends the binary group structure relationship in the traditional CRF to the ternary, and can consider the time-space relationship of the more moving target, the triple structure facilitates the optimization of the tracking algorithm and the performance improvement (4), and the traditional data association model of the multi-target tracking method based on the joint expression of the data is mostly modeled according to the difference between the data, for example, the distance between the two features is calculated; This operation is essentially a dimension reduction process that will result in a partial loss of the original data. In this paper, a multi-objective tracking algorithm based on data joint distribution modeling is presented in the traditional CRF model. The method constructs a binary potential function to characterize the correlation between the track slices, and constructs a high-order class loss function (regular term) to constrain the number of targets to be solved, and on the basis of which, a cost equation is obtained, and finally the class calibration under the CRF model is realized through the cost minimization. In which the potential function between the track pieces is modeled as the data joint distribution under two hypothetical conditions, and the reasoning process of the CRF model is completed by establishing the compatibility of the related data, the repulsion probability and the completion of the CRF model. The method is characterized in that the distribution characteristics of the sample categories stored among the tree leaf nodes of the Hough forest are utilized, and the estimation of the joint distribution probability of the data under the two assumptions is realized in the form of a non-reference. The simulation experiments carried out on a plurality of databases prove the effectiveness of the method in this paper. The proposed method is a new way to design the multi-object tracking algorithm.
【學位授予單位】：華中科技大學
【學位級別】：博士
【學位授予年份】：2016
【分類號】：TP391.41
本文編號：2493621
[Abstract]:The multi-objective tracking of video sequences is an important content in the field of computer vision, and has been widely used in the fields of national defense, video surveillance, intelligent navigation/ assistant driving, intelligent robot, behavioral analysis, video retrieval, biomedicine and so on. The purpose of the video multi-target tracking is to calibrate the motion trajectory of each target in the video sequence. However, the study of multi-objective tracking algorithm is a challenging task, which is influenced by the image quality, noise and background interference, the target appearance and the motion pattern, the uncertainty of the number of tracked targets, and the complex and variable occlusion. There is also a large number of theoretical and technical problems to be solved. In the last ten years, with the continuous improvement of the performance of the target detector, the target tracking method based on the detection has attracted wide attention and has become the current multi-target tracking method. In the method, the detection response output by the detector is taken as an input, and the detection responses belonging to the same target are connected one by one by the data association technology, and finally the motion trail of each target is obtained. The design of the association model is the key to the data association technology based on the detection, and a good correlation model should be used to integrate the observation information reflecting the intrinsic attributes of the target track as much as possible, and provide a reliable track association under the noise and the complex scene. This paper focuses on the development of the multi-object tracking field, and focuses on the correlation model and its application in the multi-objective tracking method. The main work includes the following four parts: (1) The multi-objective tracking method based on the Hough forest learning is an effective multi-target tracking method. In this paper, a multi-target tracking algorithm based on the Hough forest classifier is proposed in this paper. the method comprises the following steps of: firstly, generating a reliable short track slice through a conservative correlation algorithm; then, extracting the appearance and the motion characteristic with the discrimination performance from the trace piece set in a step-by-step processing mode, generating a training sample and constructing a Hough forest; and in the testing phase, The connection probability between the track slices is estimated by using the effective symbol information stored in the forest leaf node, and finally the path association is converted into a solution problem under the maximum posterior probability criterion (MAP). The experiment proves the effectiveness of the correlation model of the track slice based on the Hough forest: compared with the recent algorithm of some foreign peers, the method has obtained the corresponding tracking effect (2) matching the complexity of the motion scene due to the isolated response point of the occlusion reasoning model, The frequently occurring occlusion, etc., even the most advanced detector at present, also has the problems of false detection, missed detection, inaccurate detection, etc. The conservative association strategy adopted in the reliable track slice generation phase will also miss some detection responses. The above problems will result in an isolated response point that cannot be associated with any trace in the final tracking result, so that the target track gap is increased and the smoothness decreases. In this paper, a new occlusion reasoning model is proposed, which can be used to infer the occluded area and the non-occlusion area of the occlusion target. Based on this, the fusion feature description of the occlusion target is designed, and the matching strategy between the isolated response point and the target track is proposed. And the problem of the target attribution of the isolated response point is effectively solved. As a post-processing technique to fill the gap of the track, the method of this chapter is of general applicability to the tracking algorithm associated with the track slice. (3) The multi-objective tracking method based on the conditional random field (CRF) has become a hot topic in recent years based on the multi-objective tracking method of the Hough forest condition with the airport. As the core of the CRF model, the parameter estimation and state-based reasoning of the CRF model is very difficult, and the parameters are estimated conventionally by using a heuristic algorithm or a heuristic method, and the CRF reasoning process is also easy to fall into a local optimal. In this paper, Hough Forest Conditional Random Field (HFRF) is proposed in this paper. The probability of MH jump acceptance is calculated by the SW-ctrl algorithm to realize the state-based reasoning, and the probability parameters required for CRF inference are provided by the Hough forest. HFRF has embedded CRF model parameter learning and reasoning in the same framework, thus avoiding the difficulty in the traditional CRF tracking method. In addition, unlike the traditional CRF map model, HFRF additionally defines a binary indicator hidden variable for each side, extends the binary group structure relationship in the traditional CRF to the ternary, and can consider the time-space relationship of the more moving target, the triple structure facilitates the optimization of the tracking algorithm and the performance improvement (4), and the traditional data association model of the multi-target tracking method based on the joint expression of the data is mostly modeled according to the difference between the data, for example, the distance between the two features is calculated; This operation is essentially a dimension reduction process that will result in a partial loss of the original data. In this paper, a multi-objective tracking algorithm based on data joint distribution modeling is presented in the traditional CRF model. The method constructs a binary potential function to characterize the correlation between the track slices, and constructs a high-order class loss function (regular term) to constrain the number of targets to be solved, and on the basis of which, a cost equation is obtained, and finally the class calibration under the CRF model is realized through the cost minimization. In which the potential function between the track pieces is modeled as the data joint distribution under two hypothetical conditions, and the reasoning process of the CRF model is completed by establishing the compatibility of the related data, the repulsion probability and the completion of the CRF model. The method is characterized in that the distribution characteristics of the sample categories stored among the tree leaf nodes of the Hough forest are utilized, and the estimation of the joint distribution probability of the data under the two assumptions is realized in the form of a non-reference. The simulation experiments carried out on a plurality of databases prove the effectiveness of the method in this paper. The proposed method is a new way to design the multi-object tracking algorithm.
【學位授予單位】：華中科技大學
【學位級別】：博士
【學位授予年份】：2016
【分類號】：TP391.41
【相似文獻】
相關期刊論文 前10條
1 金忠;一種多目標跟蹤算法[J];南京理工大學學報(自然科學版);1985年S1期
2 龔萍;張輝;毛征;張慶龍;孔文超;;融合局部熵二維熵的空中目標跟蹤算法研究[J];國外電子測量技術;2014年01期
3 馬奔,史忠科,皮燕妮;成像目標跟蹤算法分析[J];西安電子科技大學學報;2005年03期
4 孫中森;孫俊喜;宋建中;喬雙;;一種抗遮擋的運動目標跟蹤算法[J];光學精密工程;2007年02期
5 陳愛華;孟勃;朱明;王艷華;;多模式融合的目標跟蹤算法[J];光學精密工程;2009年01期
6 牛長鋒;劉玉樹;;融合多特征的粒子濾波目標跟蹤算法[J];華中科技大學學報(自然科學版);2010年01期
7 蔡榮太;吳元昊;王明佳;吳慶祥;;視頻目標跟蹤算法綜述[J];電視技術;2010年12期
8 佟國峰;蔣昭炎;谷久宏;龐曉磊;;基于隨機蕨叢的長期目標跟蹤算法[J];東北大學學報(自然科學版);2013年01期
9 曹曉麗;李明;邢玉娟;譚萍;;幾種自動目標跟蹤算法的比較研究[J];硅谷;2013年02期
10 王魯平,李飚,胡敏露;一種基于多傳感器數據融合的目標跟蹤算法[J];紅外與激光工程;2004年02期
相關會議論文 前10條
1 徐炳吉;;一種多站聯(lián)合目標跟蹤算法[A];數學及其應用文集——中南模糊數學和系統(tǒng)分會第三屆年會論文集（上卷）[C];1995年
2 杜方芳;劉士榮;邱雪娜;;一種改進的粒子濾波目標跟蹤算法[A];PCC2009—第20屆中國過程控制會議論文集[C];2009年
3 付曉薇;方康玲;李曦;;一種基于特征的多目標跟蹤算法[A];2003年中國智能自動化會議論文集（下冊）[C];2003年
4 許偉村;趙清杰;;一種基于粒子濾波的多目標跟蹤算法[A];2011年中國智能自動化學術會議論文集（第一分冊）[C];2011年
5 李軍;張華;單梁;;一種基于Mean shift和粒子濾波的綜合目標跟蹤算法[A];2009年中國智能自動化會議論文集（第七分冊）[南京理工大學學報（增刊）][C];2009年
6 肖敬若;胡伏原;鄭江濱;張艷寧;;一種有效的多目標跟蹤算法[A];第十二屆全國信號處理學術年會（CCSP-2005）論文集[C];2005年
7 鄭黎義;陳興無;王磊;李正東;;紅外/雷達雙傳感器融合目標跟蹤算法[A];中國工程物理研究院科技年報（2005）[C];2005年
8 張震宇;王立松;;基于粒子濾波的傳感器目標跟蹤算法[A];2008年中國高校通信類院系學術研討會論文集（上冊）[C];2009年
9 王亞楠;陳杰;甘明剛;;基于差分進化的改進粒子濾波目標跟蹤算法[A];中國自動化學會控制理論專業(yè)委員會C卷[C];2011年
10 張濤;費樹岷;胡剛;;基于多特征信息自適應融合的視頻目標跟蹤算法[A];第二十九屆中國控制會議論文集[C];2010年
相關博士學位論文 前10條
1 胡子軍;基于隨機有限集的雷達多目標跟蹤算法研究[D];西安電子科技大學;2015年
2 張雷;復雜場景下實時目標跟蹤算法及實現(xiàn)技術研究[D];中國科學院研究生院(長春光學精密機械與物理研究所);2016年
3 王晶晶;復雜擁擠環(huán)境下協(xié)同視頻監(jiān)控中目標跟蹤算法研究[D];中國科學技術大學;2016年
4 田浩;基于核函數的目標跟蹤算法研究[D];長安大學;2016年
5 秦永;基于隨機有限集的目標跟蹤算法研究[D];華中科技大學;2016年
6 項俊;基于檢測的數據關聯(lián)多目標跟蹤算法研究[D];華中科技大學;2016年
7 盧建國;基于粒子濾波的視頻目標跟蹤算法研究[D];北京郵電大學;2011年
8 馮巍;分布式多視角目標跟蹤算法研究[D];復旦大學;2011年
9 王書朋;視頻目標跟蹤算法研究[D];西安電子科技大學;2009年
10 劉晴;基于區(qū)域特征的目標跟蹤算法研究[D];北京理工大學;2014年
相關碩士學位論文 前10條
1 張健;形變目標跟蹤算法的研究與實現(xiàn)[D];遼寧大學;2015年
2 張巧麗;基于LabVIEW的運動目標跟蹤算法研究與實現(xiàn)[D];陜西科技大學;2015年
3 鐘寶康;基于壓縮感知的預測目標跟蹤算法研究[D];江西理工大學;2015年
4 薛桐;基于CamShift的運動目標跟蹤算法研究[D];沈陽理工大學;2015年
5 王增宇;基于稀疏表達的目標跟蹤算法研究[D];沈陽理工大學;2015年
6 王靜;結構化的表觀模型及兩階段目標跟蹤算法研究[D];沈陽理工大學;2015年
7 葛凱蓉;自然場景下目標跟蹤算法的研究[D];山東大學;2015年
8 向偉;基于檢測的目標跟蹤算法研究[D];上海交通大學;2015年
9 單順勇;結合多示例學習和模板匹配的目標跟蹤算法研究[D];江西理工大學;2015年
10 張碧武;基于單目視覺的目標跟蹤算法的研究與實現(xiàn)[D];電子科技大學;2015年
，本文編號：2493621
本文鏈接：http://sikaile.net/shoufeilunwen/xxkjbs/2493621.html
最近更新
教材專著
