matlab图像识别英文文献,matlab图像处理-外文翻译-外文文献-英文文献-基于视觉的矿井救援...

《matlab图像处理-外文翻译-外文文献-英文文献-基于视觉的矿井救援》由会员分享，可在线阅读，更多相关《matlab图像处理-外文翻译-外文文献-英文文献-基于视觉的矿井救援(12页珍藏版)》请在人人文库网上搜索。

1、附录A英文原文Scene recognition for mine rescue robotlocalization based on visionCUI Yi-an(崔益安),CAI Zi-xing(蔡自兴),WANG Lu(王 璐)Abstract : A new scene recognitionsystem was presented based on fuzzy logic andhidden Markov model(HMM) that can be applied in mine rescue robot localization during emergencies. The 。

2、system uses monocular camera to acquire omni-directional images of the mine environment where the robot locates. By adopting center-surround differenee method, the salient local image regions are extracted from the images as n atural la ndmarks. These Ian dmarks are orga ni zed by using HMM to repre。

3、se nt the scene where the robot is, and fuzzy logic strategy is used to match the scene and Ian dmark. By this way, the localizati on problem, which is the scene recog niti on problem in the system, can be con verted into the evaluatio n problem of HMM. The con tributio ns of these skills make the s。

4、ystem have the ability to deal with cha nges in scale, 2D rotatio n and viewpoi nt.The results of experime nts also prove that the system has higher ratio of recog niti on and localizatio n in both static and dyn amic mi ne environmen ts.Key words:robot location; scene recognition;salient image; mat。

5、ching strategy; fuzzylogic; hidde n Markov model1 In troducti onSearch and rescue in disaster area in the domain of robot is a burgeoning and challenging subject1. Mine rescue robot was developed to enter mines during emerge ncies to locate possible escape routes for those trapped in side and determ。

6、 ine whether it is safe for huma n to en ter or not. Localizati on is a fun dame ntal problem in this field. Localization methods based on camera can be mainly classified into geometric, topological or hybrid ones2. With its feasibility and effectiveness, scene recognition becomes one of the importa。

7、 nt tech no logies of topological localizati on.Currently most scene recognition methods are based on global image features and have two dist inct stages: training offline and match ing on li ne.During the trainingstage, robot collects the images of the environment where itworks and processes the im。

8、ages to extract global features that represe nt the scene. Some approaches were used to an alyze the data-set of image directly and some primary features were found, such as the PCA method 3. However, the PCA method is not effective in disti nguish ing the classes of features. Ano ther type of appro。

9、ach uses appeara nee features in cludi ng color, texture and edge den sity to represe nt the image. For example, ZHOU et al4 used multidimensional histograms to describe global appearanee features. This method is simple but sensitive to scale and illumination changes. In fact, all kinds of global im。

10、age features are suffered from the cha nge of environment.LOWE 5 prese nted a SIFT method that uses similarity in varia nt descriptors formed by characteristic scale and orientationat interest points to obtain the features. Thefeatures are invariant to image scaling, translation,rotation and partial。

11、ly invariant toillumi natio n cha nges. But SIFT may gen erate 1 000 or more in terest poi nts, which may slow dow n the processor dramatically.During the matchi ng stage, n earest n eighbor strategy(NN) is widely adopted for its facility and in telligibility6. But it cannot capture the con tributi 。

12、on of in dividual feature for sce ne recog niti on.In experime nts, the NN is not good eno ugh to express thesimilarity betwee n two patter ns. Furthermore, the selected features can not represe nt the scene thoroughly according to the state-of-artpatternrecognition,which makesrecog niti on not reli。

13、able7.So in this work a new recognition system is presented, which is more reliable and effective if it is used in a complex mine environment. In this system, we improve the invarianee by extracting salient local image regions as Iandmarks to replace the whole image to deal with large cha nges in sc。

14、ale, 2D rotatio n and viewpoi nt. And the nu mber of interest points is reduced effectively, which makes the processingeasier. Fuzzyrecognitionstrategy is designedto recognize the Iandmarks in place of NN,which canstre ngthe nthe con tributi on ofin dividual featurefor sce ne recog niti on. Becauseo。

15、f itspartial in formatio n resu ming ability, hidde n Markov modelis adopted to orga nize thoseIandmarks,which can capturethe structure orrelationship among them.Soscenerecognition can be transformed to the evaluation problem of HMM, which makes recog niti on robust.2 Salie nt local image regi ons d。

16、etectio nResearches on biological visio n system in dicate that orga nism (like drosophila) ofte npays atte nti on to certa in special regi ons in the scene for their behavioral releva nee orlocal image cues while observi ng surro undings 8. These regi ons can be take n as n aturalIan dmarks to effe。

17、ctively represe nt and dist in guishdiffere nt environmen ts.In spired bythose, we use cen ter-surr ound differe nee method to detect salie nt regi ons in multi-scale image spaces. The opp onen cies of color and texture are computed to create the salie ncy map.Follow-up, sub-image cen tered at the s。

18、alie nt positi on inS is take n as the Ian dmarkregion. The size of the Iandmarkregion can be decided adaptively according to thecha nges of gradie nt orie ntati on of the local image 11.Mobile robot navigation requires that natural Iandmarks should be detected stablywhen environments change to some。

19、 extent. To validate the repeatability on Iandmark detect ion of our approach, we have done some experime nts on the cases of scale, 2D rotati on and viewpo int cha nges etc. Fig.1 shows that the door is detected for its salie ncy whe n viewpo int cha nges. More detailed an alysis and results about 。

20、scale and rotati on can be found in our previous works12.3 Scene recog niti on and localizati onDifferent from other scene recognition systems, our system doesn need training offline. In other words, our scenes are not classified in advanee. When robot wanders, sce nes captured at in tervals of fixe。

21、d time are used to build the vertex of a topological map, which represents the place where robot locates. Although the map s geometric layout isignored by the localizationsystem, it is useful for visualization and debugging13 andbeneficial to path planning. So localization means searching the best m。

22、atch of current sce ne on the map. In this paper hidde n Markov model is used to orga nize the extractedIandmarks from current scene and create the vertex of topological map for its partial in formatio n resu ming ability.Resembled by panoramic vision system, robot looks around to get omni-images.Fr。

23、omFig.1 Experime nt on viewpo int cha ngeseach image, salient local regions are detected and formed to be a sequenee, named asIandmark sequenee whose order is the same as the image sequenee. Then a hiddenMarkov model is created based on the Ian dmark seque nee in volvi ngk salie nt local imageourwer。

24、egions, which is taken as the description of the place where the robot locates. In system EVI-D70 camera has a view field of 170 Considering the overlap effect,sample en vir onment every 45to get 8 images.Let the 8 images as hidden state Si (1 4 8), the created HMM can be illustrated byFig.2. The pa。

25、rameters of HMM, aij and bjk, are achieved by learning, using Baulm-Welch algorithm14. The threshold of con verge nee is set as 0.001.As for the edge of topological map, we assig n it with dista nee in formati on betwee ntwo vertices. The dista nces can be computed accord ing to odometry readi ngs.A。

26、i5Fig.2 HMM of environmentLkeye on environment andTo locate itself on the topological map, robot must run itsextract a Iandmark sequenee L1 -Lk , then search the map for the best matched vertex(seen e). Differe nt from traditi onal probabilistic localizati on 15, in our system localizati on problem 。

27、can be con verted to the evaluati on problem of HMM. The vertex with the greatest evaluati on value, which must also be greater tha n a threshold, is take n as the best matched vertex, which in dicates the most possible place where the robot is.4 Match strategy based on fuzzy logicOne of the key iss。

28、ues in image match problem is to choose the most effective featuresor descriptors to represe nt the origi nal image. Due to robot moveme nt, those extractedIan dmark regi ons will cha nge at pixel level. So, the descriptors or features chose n should be in varia nt to some exte nt accord ing to the 。

29、cha nges of scale, rotati on and viewpo int etc.In this paper, we use 4 features commonly adopted in the community that are briefly described as follows.GO: Gradie nt orie ntati on .It has bee n proved that illu min ati on and rotati on cha ngesare likely to have less in flue nee on it5.ASM and ENT:。

30、 Angular second moment and entropy, which are two texture descriptors.H: Hue, which is used to describe the fun dame ntal in formatio n of the image.Ano ther key issue in match problem is to choose a good match strategy or algorithm.Usually nearest neighbor strategy (NN) is used to measure the simil。

31、arity between two patter ns. But we have found in the experime nts that NN can adequately exhibit the in dividual descriptor or feature s con tributio n to similarity measureme nt. As in dicated inFig.4, the in put image Fig.4(a) comes from differe nt view of Fig.4(b). But the dista nee betwee n Fig。

32、s.4(a) and (b) computed by Jefferey diverge nee is larger tha n Fig.4(c).To solve the problem, we design a new match algorithm based on fuzzy logic for exhibit ing the subtle cha nges of each features. The algorithm is described as below.And the Ian dmark in the database whose fused similarity degre。

33、e is higher tha n anyothers is take n as the best match. The match results of Figs.2(b) and (c) are dem on stratedby Fig.3. As indicated, this method can measure the similarity effectively between two patter ns.Landmark in database (index)Fig.3 Similarity computed using fuzzy strategy5 Experime nts 。

34、and an alysisThe localizati on system has bee n impleme nted on a mobile robot, which is built byour laboratory. The vision system is composed of a CCD camera and a frame-grabber IVC-4200. The resoluti on of image is set to be 400320 and the sample freque ncy is setto be 10 frames/s. The computer sy。

35、stem is composed of 1 GHz processor and 512 M memory, which is carried by the robot. Prese ntly the robot works in in door en vir onmen ts.Because HMM is adopted to represe nt and recog nize the scene, our system has theability to capture the discrim in ati on about distributi on of salie nt local i。

36、mage regi ons anddistinguish similar scenes effectively. Table 1 shows the recognition result of static en viro nments in cludi ng 5 Ian eways and a silo. 10 scenes are selected from each en viro nment and HMMs are created for each scene. Then 20 scenes are collected whe nthe robot en ters each en v。

37、iro nment subseque ntly to match the 60 HMMs above.In the table,“ truth ” means that the scene to be localized matches with the right scene(the evaluationvalue of HMM is 30% greater than the second high evaluation).“ Uncertainty” means that the evaluation value of HMM is greater than the second high。

38、evaluation under10%.“ Error match ” means that the scene to be localized matches withthe wrong scene. In the table, the ratio of error match is 0. But it is possible that the scene to be localized can match any scenes and new vertexes are created. Furthermore, the“ ratio of truth ” about silo is low。

39、er because salient cues are fewer in this kind of en viro nment.In the period of automatic explori ng, similar scenes can be comb in ed. The processcan be summarized as: when localizationsucceeds, the current Iandmark sequenee isadded to the accompanying observation sequenee of the matched vertex un。

40、-repeatedly according to their orientation (including the angle of the image from which the salient local region and the heading of the robot come). The parameters of HMM are learned aga in.Compared with the approaches using appearanee features of the whole image (Method 2, M2), our system (M1) uses。

41、 local salie nt regio ns to localize and map, which makes it have more toleranee of scale, viewpoint changes caused by robot movement and higher ratio of recog niti on and fewer amount of vertices on the topological map. So, our system has better performa nee in dyn amic environment. These can be se。

42、e n in Table 2.Lan eways 1, 2, 4, 5 are in operati on where some miners are work ing, which puzzletherobot.6 Con clusi ons1) Salie nt local image features are extracted to replace the whole image to participatein recognition, which improve the toleranee of changes in scale, 2D rotationandviewpo int 。

43、of environment image.2) Fuzzy logic is used to recog nize the local image, and emphasize the in dividual feature s contribution torecognition, which improves the reliability of Iandmarks.3) HMM is used to capture the structure or relati on ship of those local images, whichcon verts the scene recog n。

44、iti on problem into the evaluati on problem of HMM.4) The results from the above experime nts dem on strate that the mine rescue robotsee ne recog niti on system has higher ratio of recog niti on and localizati on.Future work will be focused on using HMM to deal with the uncertaintyof localizatio n.。

45、附录B 中文翻译基于视觉的矿井救援机器人场景识别CUI Yi-an(崔益安),CAI Zi-xing(蔡自兴),WANG Lu(王 璐)摘要：基于模糊逻辑和隐马尔可夫模型(HMM,论文提出了一个新的场景识别系 统，可应用于紧急情况下矿山救援机器人的定位。该系统使用单眼相机获取机器 人所处位置的全方位的矿井环境图像。 通过采用中心环绕差分法，从图像中提取 突出的位置图像区域作为自然的位置标志。 这些标志通过使用HMMT机组织起来 代表机器人坐在场景，模糊逻辑算法用来匹配场景和位置标志。通过这种方式， 定位问题，即系统的现场识别问题，可以转化为对HMM勺评价问题。这些技术贡 献使系统具有处理比率变。

46、化、二维旋转和视角变化的能力。实验结果还证明，该 系统在静态和动态矿山环境中都具有较高的识别和定位的成功率。关键字：机器人定位；场景识别；突出图像；匹配算法；模糊逻辑；隐马尔可夫模型1介绍在机器人领域搜索和救援灾区是一个新兴而富有挑战性的课题。矿井救援机器人的开发是为了在紧急情况下进入矿井为被困人员查找可能的逃生路线，并确定该线路是否安全。定位识别是这个领域的基本问题。 基于摄像头的定位可以主 要分为几何法、拓扑法或混合法。凭借其可行性和有效性，场景识别成为拓扑定 位的重要技术之一。目前，大多数场景识别方法是基于全局图像特征，有两个不同的阶段：离线培训和在线匹配。在训练阶段，机器人收集其所工作。

47、环境的图像，并处理这些图像提取出能表征该场景的全局特征。一些方法直接分析图像数据得到一些基本特征，比如PCA方法。但是，PCA方法是不能区分特征的类别。另一种方法使用外观特征包括颜 色、纹理和边缘密度来表示图像。例如，周等人用多维直方图来描述全局外观特 征。此方法简单，但对比率和光照变化敏感。事实上，各种全局图像特征，所受 来自环境变化的影响。LOWE提出了 SIFT方法，该方法利用关注点尺度和方向所形成的描述的相似性获得特 征。这些特征对于图像缩放、平移、旋转和局部光照不变是稳定的。但SIFT可能产生1 000个或更多的兴趣点，这可能使处理器大大减慢。在匹配阶段，近邻算法(NN因其简单和可行。

48、而被广泛采用。但是它并不能捕捉到个别特征对场景识别的贡献。在实验中，Nh在表达两种部分之间的相似性时效果 并不足够好。此外，所选的特征并不能彻底地按照国家模式识别标准表示场景， 这使得识别结果不可靠。因此，在这些分析中提出了一种新的识别系统，如果使用在复杂的矿井环境中它将更加可靠和有效。在这个系统中，我们通过提取突出的图像局部区域作为位置标志用 以替代整个图像，改善了信息的稳定性，从而处理比率、二维旋转和视角的变化。 兴趣点数量有效减少，这使得处理更加容易。模糊识别算法用以识别邻近位置的 位置标志，它可以增强个别特征对场景识别的作用。由于它的部分信息恢复能力， 采用隐马尔可夫模型组织这些位置标。

49、志，它可以捕捉到的结构或标志之间的关 系。因此，场景识别可以转化为对 HM评价问题，这使得识别具有鲁棒性。2局部图像区域不变形的检测生物视觉系统的研究表明，生物体(像果蝇)在观察周围环境时，经常因为他们的行为 习惯注意场景中确定的特殊区域或者局部图像信息。这些区域可以当作天然的位置标志有效地表示和区别不同环境。 受这些启示，我们利用中心环绕差分法检测多尺度图像空 间突出的区域。计算颜色和纹理的相似度用以绘制突出区域的地图。随后，以地图突出位置为中心的分图像，被定义为位置标志区域。 位置标志区域的大小可以根据该区域图像梯度方向的变化自适应决定。移动机器人的导航要求当环境有一定程度变化时自然位置标。

50、志能被稳定地检测出来。为了验证我们方法对位置标志检测的的可重复性，我们已经在图像比例、 二维旋转和视角等变化时，做了一些实验。图1表明当视角变化时因为它的突出效果大门能被检测出来。关于比 率和旋转更详细的分析和结果可以在我们以前的论文中发现。图1关于视角变化的实验3场景识别和定位与其他场景识别系统不同的是，我们的不需要离线培训。换句话说，在前进 中，我们不必对场景分类。当机器人徘徊时，在固定时间间隔内捕获的场景用于 生成拓扑地图的顶点，它表示了机器人所在位置。虽然地图的几何布局被定位系 统忽视，但它对可视化调试是有用的，并对路径规划很有益处。因此，定位即意 味在地图上搜索当前场景的最佳匹配位置。

51、。 在论文中隐马尔可夫模型是用来组织 从当前的现场提取的位置标志，并且设置拓扑图的顶点用以实现部分信息的恢复 能力。类似于全景视觉系统，机器人环顾四周，以获得全方位的图像。从每一张图像，探测突出的局部区域，并形成一个命名为“位置标志序列”的序列，其顺序和图像 序列相同。然后，在包含k个突出局部图像区域的位置标志序列的基础上，创建 一个隐马尔可夫模型当作机器人所处环境的描述。在我们的系统中EVI- D7(相目机 拥有一个土 170。视野。为了避免重复，我们的样品图像每 45。取一张，一共 得到8个图像。让这8个图像作为隐藏状态S (K i 8)，创建的HM模型可以由图3描述出 来。HMI的参数a。

52、j和bjk是通过使用Baulm - Welch算法学习获得。衔接的阈值设置 为 0.001。图2机器人所处环境的HMM至于拓扑图的边缘，我们用两个顶点之间的距离信息设置。距离可以根据里 程计的读数得到4基于模糊逻辑的匹配方法图像匹配的关键问题之一，是要选择最有效的特征或描述代表原始图像。 由于机器人的运动，那些提取位置标志的区域将在像素级内改变。 因此，选择的描 述或特征相应于比例、旋转和视角的变化应该在一定范围内保持恒定不变， 在本 文中，我们使用社会上通常采用的四个特征，简单地描述如下：GO:渐变的方向。已经证明光照和旋转的变化可能对它有更少的影响。ASM口ENT角二阶矩和熵，这是两个纹理。

53、描述符。H：色调，用来描述图像的基本信息。匹配问题的另一个关键问题是选择一个好的匹配方法或算法。通常近邻方法(神经网络)是用来测量两部分之间的相似性。 但我们在实验中发现，神经网络 不能充分展现个体描述或特征对相似性度量的作用。为了解决这个问题，我们设计了一个新的以模糊逻辑为基础的匹配算法，用以展示每个特征的微妙变化。而在数据库中的融合相似程度最高的位置标志， 可以当作最佳匹配。匹配结 果如图5所示。如上所述，该方法可以有效地测量两部分图像的相似性。5实验和分析该定位系统已经应用在我们实验室制造移动机器人。该视觉系统是由一个0 8 6 4 2 I I o o o O (/leJXUUMPrH 。

54、ytNJru jo 二口vlUMCC相机和一帧采集卡IVC-4200。图像分辨率设置为400 X 320，采样频率设置 为10帧/秒。该电脑系统是由1千兆赫处理器和512M内存，由机器人携带。目前机 器人运行于室内环境。Judgement of acceptanceFlest iiwch Judijemtznt of rejection 1(丿加%丿48121620Landmark in database (index)图3使用模糊逻辑方法计算的相似度由于采用HM的描述和识别场景，我们的系统有能力捕捉图像局部区域分布的区别，并且能有效区分类似的场景。 从每个环境选取10个场景，并为每个场景 创。

55、建隐马尔可夫模型。然后收集20个场景时为机器人随后进入每一个环境， 用以 形成60个以上HMM。在自动搜索过中，类似的场景可以连接起来。这个过程可以概括为：当定位 成功，当前的位置标志序列根据他们的方向添加到伴随观察的匹配顶点的序列(包括从突出局部区域和机器人顶部所观测的图像角度)。HMM模型的参数重复 学习。相对于使用整体图像外观特征的方法(方法 2, M2，我们的系统(M1使用 局部突出区域定位和识别地图，这使得它能有更多对变化的容许， 这些变化由机 器人的运动、高频率的识别和拓扑图上较少的顶点数量所造成。 因此，我们的系 统在动态环境中具有更好的性能。6结论1)提取突出局部图像特征，以取代整体图像参与的识别，这提高了对比例 大小的变化，二维旋转和环境图像视角的容许。2)模糊逻辑用来识别局部图像，并且强调个体特征对识别的作用，从而提高了位置标 志的可靠性3) HM模型是用来捕获图像结构或那些局部图像之间的联系，这使场景识别 问题转化为对HMI的评价问题。4)实验结果表明，该矿山救援机器人场景识别系统具有较高的识别和定位的比例。今后的工作将侧重于利用 HM的处理定位的不确定性。