本文關鍵詞： 協(xié)同跟蹤 無線傳感器網絡 目標檢測 行為識別 特征提取　出處：《西安電子科技大學》2016年博士論文　論文類型：學位論文

【摘要】：隨著嵌入式技術、通信技術和計算機視覺技術的高速發(fā)展,無線傳感器網絡以其先進的理念和廣闊的應用前景日益受到學術界的關注,相關技術也成為當前國際上新興的研究熱點之一。運動目標的協(xié)同跟蹤作為無線傳感器網絡的一種典型應用一直備受關注,但目前的研究大多針對高空飛行目標,對日常生活中的運動目標例如人體的跟蹤則少有涉及。智能視頻監(jiān)控技術是在計算機視覺和圖像處理技術上,結合其它相關技術和理論發(fā)展起來的一個較新的研究領域,旨在利用計算機或智能處理單元的數(shù)據(jù)分析能力,自動實現(xiàn)視頻場景中靜態(tài)和動態(tài)實物的感知、描述以及分析,滿足日常生產、生活中智能安防、智能交通以及城市智慧化建設需要。因此將無線傳感器網絡應用在智能視頻監(jiān)控系統(tǒng)中,能夠實現(xiàn)無線傳感器網絡對可疑目標的分析和協(xié)同跟蹤,具有一定的研究及應用價值�；跓o線傳感器網絡的運動目標協(xié)同跟蹤涉及的知識面較為廣泛,按照工作流程主要包含的技術問題和關鍵步驟有：網絡資源的優(yōu)化部署、運動目標檢測和跟蹤、目標行為分析、特征提取和匹配、協(xié)同跟蹤算法等。雖然現(xiàn)有的視頻圖像分析技術能夠解決在某些應用場景下的以上問題,但對于無線傳感器網絡自身的局限性,例如無線傳感器節(jié)點對能耗敏感和運算能力有限等問題并不適用,因此本文針對以上提及的關鍵技術和難點展開研究,具體內容如下：(1)針對無線傳感器網絡優(yōu)化部署的效率問題,提出了建立在職能劃分基礎上的網絡優(yōu)化部署算法。由于智能監(jiān)控領域對目標的協(xié)同跟蹤,往往是針對可疑目標來進行的,因此提出了在網絡初始化階段首先區(qū)分無線傳感器節(jié)點的職能,將網絡監(jiān)控點細分為兩大類：行為識別監(jiān)控點和協(xié)同跟蹤監(jiān)控點。然后針對系統(tǒng)存在部分可移動協(xié)同跟蹤監(jiān)控點的情況,設置行為識別監(jiān)控點為初始化聚類中心,采用動態(tài)模糊聚類算法進行網絡優(yōu)化部署,而對只存在靜態(tài)監(jiān)控點的系統(tǒng)采用改進的粒子群優(yōu)化算法進行網絡部署。將監(jiān)控點的職能劃分和優(yōu)化部署算法相結合的方案,有利于充分發(fā)揮無線傳感器網絡的固有優(yōu)勢,為可疑目標的確定和特征提取打下基礎。(2)針對可疑目標的篩選問題以及無線傳感器節(jié)點的局限性,提出了一種適用于無線傳感器網絡的運動人體行為識別法。通過行為識別實現(xiàn)可疑目標的定位可分以下幾步：運動目標檢測、跟蹤和行為識別。為了克服運動目標檢測中遇到的場景多變的干擾和無線傳感器節(jié)點運算能力的局限性,采用背景減除法和局部廣義霍夫投票相結合的方法進行運動檢測,能夠較為完整地提取出運動目標區(qū)域。而運動目標的跟蹤采用基于檢測的方法來實現(xiàn),通過持續(xù)的運動目標檢測,達到單節(jié)點跟蹤的目的。最后對于可疑目標的判定問題,提出了建立行為模板庫,通過運動目標輪廓小波矩和速度小波矩的提取,結合行為庫的模板匹配法來判斷目標的行為,若行為異常則確定為待協(xié)同跟蹤的目標。僅對可疑目標進行協(xié)同跟蹤,更加符合實際系統(tǒng)的應用需求。(3)針對不同監(jiān)控點環(huán)境差異對運動目標特征提取的影響,而復雜的特征提取算法不適用于無線傳感器節(jié)點的實際問題,提出了一種多角度數(shù)據(jù)融合的可疑目標特征提取與匹配算法。首先利用無線傳感器網絡中監(jiān)控點存在重復監(jiān)控區(qū)域覆蓋的特性,不同角度的監(jiān)控點將可疑目標輪廓外接矩形內部的像素區(qū)域進行超像素分割,對形成的有限個超像素區(qū)域進行顏色特征表達,然后將多角度獲得的超像素區(qū)域顏色特征進行數(shù)據(jù)融合,得到可疑目標的特征表達。在協(xié)同跟蹤監(jiān)控點進行特征匹配時,對當前運動目標進行類似的特征提取,再采用兩層匹配法進行特征匹配,由匹配結果判斷當前運動目標是否為協(xié)同跟蹤目標。該方法能夠降低不同場景下對同一可疑目標特征提取的誤差,提高特征匹配精度。(4)針對無線傳感器網絡的能耗問題,提出了一種建立在休眠與喚醒機制上的幾何監(jiān)控區(qū)域近似和軌跡預測算法。該算法默認網絡中的行為識別監(jiān)控點始終處于工作狀態(tài),而協(xié)同跟蹤監(jiān)控點處于休眠狀態(tài),通過對可疑目標運動軌跡的預測,由行為識別監(jiān)控點發(fā)送命令將涉及協(xié)同跟蹤的監(jiān)控點喚醒。此外,針對運算能力問題,尤其是行為識別監(jiān)控點多目標行為識別和子網管理的運算壓力問題,提出了一種基于DOT模型的并行計算思路,最后建立了協(xié)同跟蹤系統(tǒng)的能耗模型,并通過原型系統(tǒng)的實驗和性能仿真實驗,結合相似算法的數(shù)據(jù)對比,說明了本文所研究的協(xié)同跟蹤算法具有一定的先進性。
[Abstract]:With the rapid development of embedded technology, communication technology and computer vision technology, wireless sensor network with its advanced concept and broad application prospect has attracted the attention of academia, the related technology has also become the new research focus. Moving target collaborative tracking as a typical application of wireless sensor networks has attracted a lot of attention but, most of the current research on high flying target, the moving target in daily life such as the human body tracking are less involved. Intelligent video surveillance technology in computer vision and image processing technology, combined with other related technologies and theories developed in a relatively new field of study, aims to analyze the ability to use a computer or intelligent data processing unit, automatic realization of static and dynamic physical perception of the video scene, description and analysis, to meet the daily Production, intelligent life, intelligent transportation and intelligent city construction. So the application of the wireless sensor network in intelligent video surveillance system, can realize the wireless sensor network for suspicious target analysis and collaborative tracking, has a certain value of research and application of wireless sensor network. The moving target tracking involves more collaborative knowledge based on extensive, according to the technical problems and key steps of work process mainly include: optimizing the deployment of cyber source, moving target detection and tracking, target behavior analysis, feature extraction and matching, collaborative tracking algorithm. Although the existing video image analysis technology can solve the above problems in some application scenarios, but for limitations the wireless sensor network, such as wireless sensor nodes for sensitive and operational problems such as limited energy consumption and not applicable, Therefore, aiming at the key technology and difficulty of the above mentioned research, the specific contents are as follows: (1) aiming at the efficiency problem of optimal deployment of wireless sensor network, proposed the establishment of functional network deployment optimization algorithm based on the division of the field of intelligent monitoring. Because of synergistic tracking for the target is often carried out according to the suspicious target, so put forward in the network initialization phase we distinguish wireless sensor node functions, the network monitoring points is subdivided into two categories: behavior identification monitoring point and monitoring points. Then according to the collaborative tracking system is part of mobile collaborative tracking and monitoring points, set up monitoring points for behavior recognition to initialize cluster centers by dynamic network optimization deployment fuzzy clustering algorithm, and the system exists only static monitoring points based on Improved Particle Swarm Optimization Algorithm for network deployment monitoring. The combination of function and optimization deployment algorithm point scheme, is conducive to give full play to the inherent advantages of wireless sensor network, and determine the characteristics of suspicious target extraction basis. (2) screening of suspicious targets and to solve the problem of wireless sensor node the limitations of human motion recognition method is proposed for the wireless sensor network. Locate the suspicious target through behavior recognition can be divided into the following steps: moving target detection, tracking and behavior recognition. In order to overcome the interference encountered in the moving target detection and scene changing wireless sensor nodes the limitations of the method of using the method of background subtraction and local generalized Hof voting combination motion detection, can accurately extract the moving target area. While tracking the moving target detection method based on the achieved through continued The moving target detection, to achieve the purpose of tracking the single node. Finally the suspicious target decision problem, proposed the establishment of a behavior template library, by extracting the contour of the moving target speed of wavelet moment and wavelet moment matching method, to determine the target binding behavior library template, if the abnormal behavior is determined for collaborative target tracking. Cooperative tracking of suspicious targets, more in line with the application of the actual needs of the system. (3) the effects of different environmental monitoring points difference of moving target feature extraction, feature extraction algorithm for complex practical problems are not suitable for wireless sensor nodes, the proposed algorithm extracting and matching the suspicious target feature a multi angle data fusion the first use of monitoring points. The wireless sensor network has characteristics of repetitive coverage of monitoring area, monitoring will be suspicious object contour from different angles of internal external rectangle The pixel area of the pixel segmentation, the formation of a finite super pixel region color feature representation, and then the super pixel region color multi angle characteristics obtained by data fusion, feature expression of suspicious targets. In the collaborative tracking and monitoring point feature matching, the moving target is similar to feature extraction. The two layer matching method for feature matching, the matching results determine whether the current target for collaborative target tracking. This method can reduce the error of the same suspicious target feature extraction in different scenarios, improve the feature matching accuracy. (4) to solve the problem of power consumption in wireless sensor networks, a set up in dormancy and wake up regional monitoring mechanism on the geometric approximation and prediction algorithms. The algorithm is the default behavior identification monitoring point in the network is always in a working state and cooperative tracking supervision The control points in a dormant state, the prediction of the suspicious target track, by sending behavior recognition monitoring point command will involve monitoring point tracking cooperative wake up. In addition, according to the operation ability, especially the operation pressure monitoring point multi object behavior recognition behavior recognition and sub network management, this paper presents a calculation method of parallel based on the DOT model, the energy consumption model of cooperative tracking system is established, and through the simulation experiment and the performance of the prototype system, compared with the similarity algorithm, the cooperative tracking algorithm in this paper is advanced.

【學位授予單位】：西安電子科技大學
【學位級別】：博士
【學位授予年份】：2016
【分類號】：TP212.9;TN929.5

【參考文獻】

相關期刊論文 前1條

1 姜倫;丁華福;;關于模糊C-均值(FCM)聚類算法的改進[J];計算機與數(shù)字工程;2010年02期

，

本文編號：1531591