外文文獻(xiàn)翻譯——基于激光測距儀的行人跟蹤

1、<p><b>　　中文2862字</b></p><p>　　Active Pedestrian Following Using Laser Range Finder</p><p>　　1. INTRODUCTION</p><p>　　The ability of robots to track and follow movin

2、g targets is essential to many real life applications such as museum guidance, office or library assistance. On top of being able to track the pedestrians, one aspect of human-robot interaction is robot’s ability to foll

3、ow a pedestrian target in an indoor environment. There are various scenarios where the robot can be given instructions such as holding books in a library or carrying groceries at a store while following the pedestrian ta

4、rget.</p><p>　　The key components of moving target following technique include Simultaneous Localization and Mapping(SLAM), Detecting and Tracking Moving Objects (DATMO),and motion planning. More often than

5、not, the robots are required to operate in dynamic environments where there are multiple pedestrians and obstacles in the surroundings.Consequently, tracking and following a specific target pedestrian become much more ch

6、allenging. In other words, the following behaviors must be robust enough to deal with co</p><p>　　When designing the following algorithm, one intuitive approach is to set the target location as the destinati

7、on for the robot. However, this approach can easily lead to losing the target because it does not react to the target’s motion nor consider the visibility problem (since the target may be occluded by obstacles and become

8、 invisible). For achieving robust target following and tracking, the robot should have the intelligent to predict target motion and gather observations actively.</p><p>　　In this paper, we propose a moving t

9、arget following planner which is able to manage obstacle avoidance and target visibility problems. Experimental results are shown to compare the intuitive approach with our approach and prove the importance of active inf

10、ormation gathering in planning.This paper is organized as follow: Section II discusses related works of DATMO and planning algorithms. Section III describes our DATMO system and introduces our target following planner. L

11、astly, Section IV illust</p><p>　　2. RELATED WORKS</p><p>　　There are various approaches to detect and track moving objects such as building static and dynamic occupancy grid maps proposed by Wo

12、lf & Sukhatme [1], finding local minima in the laser scan as in Horiuchi et al.’s work [2] or using machine learning methods in Spinello et al.’s work [3]. Most of DATMO approaches assume that the robot is stationary

13、 or has perfect odometry. When tracking moving objects using mobile robots, it has been proven in Wang et al.’s work [4], that SLAM and DATMO can be d</p><p>　　and dynamic parts. In our work, we implement a

14、DATMO algorithm which is similar to the one in Montesano et al.’s work [5]. A scan matching method is used to correct robot odometry and moving points are detected by maintaining a local occupancy grid map. Moreover, Ext

15、ended Kalman Filter(EKF) is applied to track the moving objects.</p><p>　　This paper aims to solve moving target following problem with the existence of obstacles. For navigating in static environments, ther

16、e are many successful works such as Fox et al. [6], Ulrich & Borenstein [7] Minguez & Montano [8], Seder & Petrovi’c [9]. However, those methods are designed to reach a fixed goal and assume that the environm

17、ent and robot states are fully-observable. Applying traditional obstacle avoidance algorithms on the target following task can fail easily because a moving tar</p><p>　　In this paper, we propose a motion pla

18、nner for moving target following. The planner uses an extension of dynamic window approach propsed by Chou & Lian [14] to find collision-free velocities and choose a proper velocity using heuristic search. Cost funct

19、ions are designed for minimizing the distance between robot and target and maximize the possibility that the robot can keep observing the target in a fixed time horizon. Additionally, we apply the concept in nearness dia

20、gram algorithm such as the</p><p>　　3. MOVING TARGET FOLLOWING AND OBSTACLE AVOIDANCE</p><p>　　It is essential that our DATMO system is able to track the moving target, pedestrian in this case,

21、with great accuracy.The more precise pedestrian location acquired, the better the robot performs when following the target.</p><p>　　Detecting and Tracking Moving Objects</p><p>　　For detecting

22、moving points in the laser data, we adopt the concept of occupancy grid map proposed by Wolf &Sukhatme [1]. A local occupancy grid map is aintained and used to differentiate the moving points. For robot localization

23、, a scan matching technique called Iterated Closest Point (ICP) is used to acquire robot pose. The moving points are filtered out of the data prior to scan matching in order to maintain the pose accuracy. The detected mo

24、ving points are segmented into numerous clusters a</p><p>　　B. Following Method </p><p>　　Our goal is to accomplish moving target following with the existence of obstacles. A reasonable solution

25、 is to see it as a path planning problem and use obstacle avoidance algorithms to find collision-free actions. In this paper, we apply our previous work, DWA*, Chou et al. [10] as the obstacle avoidance algorithm.</p&

26、gt;<p>　　The procedure of DWA* is shown in Fig. 1 (a), it is a trajectory-rollout algorithm. The right side of Fig. 1 (a) shows the procedure for computing proper motion commands. First,the environment information

27、 is realized as interval configuration for faster processing. Each interval value represents the maximum distance can be raveled by the robot on a certain circular trajectory. Second, the intervals are clustered as navig

28、able areas. Third, for each area, a candidate velocity is determined accordi</p><p>　　In our work, we used two different methods for integrating DATMO and DWA*: pseudo goal method and trajectory optimization

29、 method. Pseudo-goal method is a more intuitive approach which does not consider the target velocity when following. Another approach is trajectory-optimization where a trajectory-rollout controller is used to approximat

30、e a predicted target trajectory and maximize target visibility.Their performances will be shown and compared in the next section.</p><p>　　Pseudo-goal method</p><p>　　When implementing people fo

31、llowing algorithm, an intuitive approach is setting the present location of the moving target as the goal of navigation algorithm. Consider the tracked target at an angle with respect to the robot, DWA* algorithm will ge

32、nerate an angular and translational velocity which produces an arc-like trajectory. If the goal is within a close proximity of the robot, the angular velocity will be small and the rc-like route often results in an indir

33、ect or detour path for the robot </p><p>　　In the pseudo-goal method, the space in front of the robot is divided into 7 trajectories at 35_ , 55_ , 70_ , 90_ ,110_ , 130_ , 145_ as shown in Fig. 3. A pseudo

34、goal is set 3 times the original distance between the goal and robot along each trajectory. After acquiring the pedestrian location (red circle in Fig. 3), the pseudo goal with trajectory closest to the pedestrian locati

35、on is then selected. By setting the goal further away from the robot, it remedies the issue of small angular velocit</p><p>　　Another problem is the limited sensing ability. In thispaper, the algorithm is im

36、plemented on a mobile robot with a 180-degree POV LRF. Here we added an “information gathering mode.” When the moving target enters the“dangerous zone,” (0~35 degree and 145~180 degree in the robot coordinate) the robot

37、will stop and turn to the target direction rapidly. Adding this mode greatly lowers the risk of losing the target.</p><p>　　Pseudo-goal method is straight-forward to implement and performs well when the targ

38、et velocity is nearly constant.However, when the target velocity changes, the pseudo-goal trajectory selected may also alter from one to another and causes the robot to change its moving direction drastically.</p>

39、<p>　　Trajectory optimization method</p><p>　　Though the pseudo-goal method is very simple to implement, it can easily lose track of the target because it does not take the robot’s limited observabilit

40、y into consideration.Another drawback is that the pseudo-goal method can’t react quickly to the change of target velocity. A better method is to plan according to a reference trajectory but not a fixed location. In this

41、way, target velocity can be considered and the robot observability can be considered by adding terms to the cost function.</p><p><b>　　外文翻譯（中文）</b></p><p>　　基于激光測距儀的行人跟蹤</p>&

42、lt;p><b>　　一、引言</b></p><p>　　在現(xiàn)實生活中機器人跟蹤運動目標(biāo)的能力得到廣泛的應(yīng)用，如博物館導(dǎo)引，辦公室或者圖書館幫助。除了能跟蹤行人，仿人機器人的另一個方面就是在室內(nèi)環(huán)境下跟蹤行人的能力。有很多種情況，機器人可以可發(fā)出指令，如當(dāng)跟蹤行人目標(biāo)時可以保護(hù)圖書館的書籍或商店里的可攜帶物品。</p><p>　　運動目標(biāo)關(guān)鍵部位的跟蹤技術(shù)包

43、括即時定位與地圖構(gòu)建（SLAM），運動目標(biāo)的檢測和跟蹤（DATMO），和運動規(guī)劃。通常機器人需要在周圍有很多行人和障礙物的動態(tài)的環(huán)境下進(jìn)行操作。因此，跟蹤一個特定的行人目標(biāo)變得更具挑戰(zhàn)性的。換句話說，跟蹤行為必須強大到足以處理在經(jīng)常擁堵的環(huán)境中躲避障礙物。</p><p>　　當(dāng)設(shè)計一個跟蹤算法時，一個直觀的方法是將目標(biāo)位置作為機器人目標(biāo)。然而，這種方法很容易導(dǎo)致失去目標(biāo)，因為它不會對目標(biāo)的運動做出反應(yīng)也沒有考慮

44、可見性的問題（因為目標(biāo)可能由于障礙物的遮擋變成無形的）。為實現(xiàn)強大的目標(biāo)追尋和跟蹤，機器人應(yīng)該有智能預(yù)測目標(biāo)的運動和聚集的觀察能力。</p><p>　　在本文中，我們提出了一個能夠處理避障和目標(biāo)可見性問題的移動的目標(biāo)跟蹤方案。實驗結(jié)果顯示相比直觀的方法，我們的方法證明了活動信息收集的重要性。本文的結(jié)構(gòu)如下：第二節(jié)討論相關(guān)DATMO和規(guī)劃算法的問題。第三節(jié)介紹了我國DATMO系統(tǒng)以及我們的目標(biāo)和下一步的計劃。最后

45、，第四部分闡述了實驗結(jié)果。</p><p><b>　　二、相關(guān)作品</b></p><p>　　有很多種不同的方法來探測和跟蹤移動對象，如Wolf & Sukhatme提出的建立靜態(tài)和動態(tài)網(wǎng)格地圖法，利用激光掃描在很小的地方尋找就像Horiuchi等人的工作或在Spinello等人的工作使用的機器學(xué)習(xí)方法。大多數(shù)DATMO方法假定機器人是固定的或具有完善的里

46、程表。當(dāng)使用移動機器人跟蹤運動物體時，它已被證明在王等人的工作。如果測量可分為靜態(tài)和動態(tài)的部分SLAM和DATMO可以同時使用。在我們的工作中，我們執(zhí)行的DATMO算法類似于在蒙特等人使用的一個。掃描匹配方法就是對機器人里程計和本地網(wǎng)格地圖中移動的點是檢測。此外，擴(kuò)展卡爾曼濾波（EKF）被施加到跟蹤移動的物體。</p><p>　　本文旨在解決移動目標(biāo)伴隨著障礙物存在時的跟蹤問題。導(dǎo)航在靜態(tài)的環(huán)境中，有許多成功的

47、例子。Ulrich & Borenstein Minguez & Montano,Seder & Petrovi’c。然而，假定環(huán)境和機器人的狀態(tài)完全可觀，這些方法設(shè)計達(dá)成既定的目標(biāo)，。應(yīng)用傳統(tǒng)的避障算法對目標(biāo)的跟蹤任務(wù)很容易失敗。因為一個移動的目標(biāo)可以改變其速度和移動方向以及目標(biāo)可以隨時被障礙遮擋。對于不完善計劃，馬爾可夫決策過程（POMDP）提供了一個總體框架。然而，利用POMDP解決最優(yōu)策略通常是非常復(fù)雜

48、的，因為它的計算規(guī)劃已經(jīng)超出了傳統(tǒng)的空間（通常是n-1維的狀態(tài)的問題。）。在實際問題中，大多數(shù)作品是以減少空間的維數(shù)使用s。如PBVI等人，AMDP Roy等人和MOMDP Sylvie等人。這些方法比原來的POMDP要快得多，但其計算復(fù)雜度仍然太高而不能做實時規(guī)劃。我們在本文的方法更像是一個最優(yōu)而且快速的方法：假設(shè)機器人總是接受最可能的可觀察到的路徑，并且規(guī)劃一個可以達(dá)到的目標(biāo)和減少不確定性特定的狀態(tài)的路徑。例如，在Prentice和

49、Roy中，機器人的目的是在最低的不確定性情況下到達(dá)目標(biāo)的位置，</p><p>　　在本文中，我們提出了一個運動規(guī)劃的移動目標(biāo)跟蹤。策劃者采用Chou&Lian提出的動態(tài)擴(kuò)展窗口的方法找到無碰撞速度和選擇合適的速度使用啟發(fā)式搜索。成本函數(shù)被設(shè)計為機器人與目標(biāo)之間的距離最小化和，可以使機器人在固定的時間范圍觀察目標(biāo)的一種最大化的可能性。此外，我們應(yīng)用的概念接近圖算法如Minguez和Montanos用于計算

50、一個更好的估計機器人與目標(biāo)之間的距離，從而實現(xiàn)光滑，快捷的性能。</p><p>　　三、運動目標(biāo)跟蹤和避障</p><p>　　我們DATMO系統(tǒng)的本質(zhì)是能夠精準(zhǔn)的跟蹤在這種情況下的行人運動的目標(biāo)。行人的位置獲得的越精確， 機器人跟蹤目標(biāo)越好。</p><p>　　A 運動目標(biāo)的檢測與跟蹤</p><p>　　在激光數(shù)據(jù)下檢測移動的點，我

51、們采用由Wolf 和Sukhatme提出的柵格地圖法。用本地網(wǎng)格地圖來區(qū)分移動點。機器人定位，一個稱為迭代最近點（ICP）的掃描匹配技術(shù)用于獲取機器人姿勢。移動的點被從之前的數(shù)據(jù)掃描過濾匹配以保持位姿精度。檢測到的移動點分割成許多集群和確定它們是否屬于一個行人使用的要素諸如運動速度，位置大小，尺寸。最后，每個行人跟蹤采用擴(kuò)展卡爾曼濾波器（EKF）解決了閉塞的問題，同時提供我們計算在實時性能上的優(yōu)勢。</p><p&g

52、t;<b>　　B 跟蹤算法</b></p><p>　　我們的目標(biāo)是實現(xiàn)在有障礙物存在情況下的運動目標(biāo)跟蹤。一個合理的解決辦法是把它作為路徑規(guī)劃問題，使用避障算法找到無碰撞的行為。在本文中，我們應(yīng)用DWA *，Chou等前人的方法作為避障算法。</p><p>　　DWA的過程如圖1(a)所示，它是一個軌道鋪設(shè)算法。圖1(a)的右側(cè)顯示用于適當(dāng)?shù)倪\動命令的程序。首

53、先,這樣的環(huán)境信息實現(xiàn)更快的處理過程。每個區(qū)間值代表可由機器人在一定的圓軌跡上行進(jìn)的最大距離。第二，這個區(qū)間應(yīng)在適航區(qū)域內(nèi)。第三，各地區(qū)，一個機器人速度是根據(jù)目標(biāo)函數(shù)的確定。為每個候選人的速度，一個新的機器人的位置被計算為在一個軌跡樹保存的新的節(jié)點。然后，一個最小估計值節(jié)點將被提取為生成新的節(jié)點的基礎(chǔ)節(jié)點。重復(fù)這一過程直到目標(biāo)位置擴(kuò)大或樹的深度達(dá)到一定的確切值。樹擴(kuò)展停止后，最深的節(jié)點被確定為時間的目標(biāo)，機器人的時間目標(biāo)的當(dāng)前速度被選擇

54、出來。</p><p>　　在我們的工作中，我們使用了兩種不同的方法整合DATMO和DWA *：偽目標(biāo)法及軌跡優(yōu)化法。偽目標(biāo)法是一種更直觀的方法，不考慮當(dāng)前目標(biāo)的速度。另一種方法是軌跡優(yōu)化法，是用軌跡控制器近似的預(yù)測目標(biāo)的運動軌跡和最大限度地提高目標(biāo)能見度。他們的運行將在下部分進(jìn)行展示和比較。</p><p><b>　　1）偽目標(biāo)法</b></p>&

55、lt;p>　　當(dāng)執(zhí)行行人跟蹤算法時，一個直觀的方法是設(shè)置移動目標(biāo)的當(dāng)前位置作為導(dǎo)航算法的目標(biāo)。考慮跟蹤目標(biāo)相對于機器人的一個角，DWA *算法將產(chǎn)生軌跡弧。如果目標(biāo)是在接近機器人，角速度會很小和圓弧形路線的往往致使機器人在間接的或迂回路徑上達(dá)到要達(dá)到的目標(biāo)。相反，如果目標(biāo)是在相同的角度進(jìn)一步遠(yuǎn)離機器人，任何的目標(biāo)運動將導(dǎo)致一個比當(dāng)目標(biāo)接近機器人更大的位移變化。因此，角速度會增加響應(yīng)目標(biāo)位置的實際變化。因此，到達(dá)目標(biāo)的軌跡更加直接了

56、。</p><p>　　在偽目標(biāo)法中，在機器人面前的空間被7 條軌跡分開，35，55，70，90，110，130，145，如圖3所示。偽目標(biāo)沿著每個軌跡把目標(biāo)和機器人之間的距離設(shè)定為原來3倍。獲取行人的位置后（圖3紅圈），偽目標(biāo)軌跡接近行人位置，然后選擇出來。按設(shè)定的比機器人更遠(yuǎn)的目標(biāo)，彌補了小的角速度問題速，同時提供更直接的路徑到達(dá)目標(biāo)。</p><p>　　另一個問題是有限的感知能力。

57、在本文里，是一個實現(xiàn)移動機器人180度的POV LRF。在這里，我們增加了一個“信息采集模式”。當(dāng)移動目標(biāo)進(jìn)入“危險地帶，”（在機器人坐標(biāo)0 ~ 35度和145~180度）機器人將停止前進(jìn)，并向目標(biāo)相反方向迅速返回。加入這種模式大大降低了失去目標(biāo)的風(fēng)險。</p><p>　　偽目標(biāo)法是實施的直線前進(jìn)，運行時的目標(biāo)速度幾乎是恒定的。然而，當(dāng)目標(biāo)速度的變化時，假目標(biāo)軌跡的選擇也可以從一個改變到另一個使機器人明顯改變其

58、運動方向。</p><p><b>　　2）軌跡優(yōu)化方法</b></p><p>　　雖然偽目標(biāo)法實現(xiàn)起來非常簡單，但是很容易失去跟蹤目標(biāo)。因為它不對機器人的有限的可觀性進(jìn)行考慮。另一個缺點是，偽目標(biāo)的方法不能快速反應(yīng)目標(biāo)速度的變化。一個更好的方法是根據(jù)一個參考軌跡規(guī)劃軌跡，但不是一個固定位置。以這種方式，可以考慮目標(biāo)速度也可以通過添加函數(shù)考慮機器人的可觀測性。<