本文關(guān)鍵詞：長期運(yùn)行移動機(jī)器人的定位與地圖構(gòu)建　出處：《浙江大學(xué)》2016年博士論文　論文類型：學(xué)位論文

  更多相關(guān)文章： 同時定位與地圖構(gòu)建 長期運(yùn)行 動態(tài)環(huán)境 概率模型

【摘要】：移動機(jī)器人領(lǐng)域的發(fā)展有效提升了機(jī)器人的自主性,使機(jī)器人具備更大的工作范圍,更低的環(huán)境部署要求,以及更靈活豐富的任務(wù)執(zhí)行等,對于機(jī)器人的進(jìn)一步應(yīng)用推廣有重要的意義。在移動機(jī)器人技術(shù)方面,最首要的任務(wù)是回答“我在哪里”,即定位問題。目前,最常見的方式是機(jī)器人先對環(huán)境進(jìn)行地圖構(gòu)建,然后利用構(gòu)建的地圖進(jìn)行定位。其中,地圖構(gòu)建依賴同時定位與構(gòu)圖(SLAM)系統(tǒng)。然而,這種方式對于長期運(yùn)行案例,比如倉儲機(jī)器人、巡檢機(jī)器人等不再可行。因為SLAM得到的地圖在定位期間可能已經(jīng)過期,甚至SLAM期間就已經(jīng)過期。針對這個問題,本文提出一種支持機(jī)器人長期運(yùn)行的多階段SLAM方法,思路是通過將長期運(yùn)行SLAM問題轉(zhuǎn)化為多階段SLAM問題,假設(shè)階段內(nèi)環(huán)境靜止,而階段間環(huán)境變化,從而給出一種多階段SLAM方案,使得移動機(jī)器人能夠適應(yīng)環(huán)境變化,構(gòu)建具有時效性的地圖,最終具有長期定位能力,并且計算復(fù)雜度可接受。具體包括了三方面五個創(chuàng)新點的內(nèi)容：(1)信息層面從冗余到精簡,本文從兩張地圖之間的Kullback-Leibler距離出發(fā),導(dǎo)出了衡量構(gòu)建兩張地圖的位姿集合間的量化差距,基于這個差距提出了圖模型位姿節(jié)點的修剪算法。在此之上,本文又提出位姿修剪后,在圖模型中生成新的稀疏因子的方法,保留一部分節(jié)點用于維持圖模型幾何關(guān)系的信息。實現(xiàn)通過控制節(jié)點使圖模型保持與地圖相關(guān)的復(fù)雜度,而非原來的軌跡長度；(2)觀測層面從靜態(tài)到動態(tài),本文從位姿估計出發(fā),利用概率模型將問題轉(zhuǎn)化為傳感器類型無關(guān)的一般化概率推斷和參數(shù)估計問題,并發(fā)現(xiàn)該框架能夠囊括許多經(jīng)典算法,并且借助該模型提出了多傳感器的融合框架。在此一般化模型的基礎(chǔ)上,本文又將環(huán)境的動態(tài)融合到模型的變量中,使動態(tài)環(huán)境檢測變?yōu)槎鄠鞲衅飨碌母怕誓Ｐ屯评韱栴},提升了計算效率和準(zhǔn)確率。(3)框架層面從單次到多次,本文通過將單次SLAM轉(zhuǎn)化為多次SLAM,并在每次SLAM之間設(shè)立包含基于修剪的冗余性處理和基于動態(tài)檢測的時效性處理的圖模型控制模塊,使SLAM可以長期運(yùn)行,僅通過損失少量精度,就可以獲得地圖有界前提下的常數(shù)復(fù)雜度。最終,通過在包含傳感器,區(qū)域,環(huán)境類型等多個變量的多個數(shù)據(jù)集上實驗,證明了所提出的系統(tǒng)的可行性及誤差非累積特性,為長期運(yùn)行機(jī)器人的SLAM問題提供了初步的理論和實踐結(jié)論。
[Abstract]:The development in the field of mobile robot can effectively enhance the autonomy of the robot, the robot has a larger working range, lower environment deployment requirements, more flexible and rich implementation, has an important significance for further application of the robot. The mobile robot technology, the most important task is to answer "I where, namely location problem. At present, the most common way is to first robot environment map building, and then use the built map positioning. The map is built upon simultaneous localization and mapping (SLAM) system. However, this way for a long running case, such as warehousing robot inspection robot is no longer feasible. Because the SLAM map may have expired during location, even during the period of SLAM has expired. To solve this problem, this paper suggests that the long-term operation of a human machine support Multi stage SLAM method, the ideas through the long run of SLAM problem is transformed into a multi stage SLAM environment problem, stationary hypothesis stage, and the stage of environmental change, which is a multi stage SLAM scheme, so that the mobile robot can adapt to the change of the environment, build a map of the final effect, with long-term capacity, complex the degree of acceptance and calculation. Including three aspects five innovation contents: (1) from the aspect of information redundancy to streamline, the map distance between two Kullback-Leibler of derived measure was constructed to quantify gap pose two maps between the sets, the pruning algorithm of graph nodes pose model based on the gap. On this basis, this paper proposes the pose after pruning method to generate a new factor in sparse graph model, some nodes reserved for maintaining the geometric relations of letter graph model Information. Through the control nodes to maintain graph model and map related complexity, rather than the original track length; (2) the observation level from static to dynamic, from the pose estimation of the problem is transformed into a general probabilistic sensor type independent inference and parameter estimation problem using the probabilistic model, and found that the framework can include many classic algorithms, and with the help of the model proposed. Multi sensor fusion framework based on this general model, this paper will environment dynamic fusion to the model variables, the dynamic environment detection to the probability model of reasoning problems under multi sensor, improve the computation efficiency and accuracy. (3) the framework level from single to multiple, the single SLAM into multiple SLAM, and in every SLAM set up based on inclusion redundancy processing and pruning based on timeliness of dynamic detection The graph model control module, so that SLAM can run only by a small amount of long-term, loss of accuracy, you can get the map with a constant circle under the premise of complexity. Finally, through the included sensor, area, multi variable environment types on the dataset experiment, feasibility and error of the proposed system the non accumulation characteristics, provides a theoretical and practical conclusions SLAM problem for the long-term operation of the robot.

【學(xué)位授予單位】：浙江大學(xué)
【學(xué)位級別】：博士
【學(xué)位授予年份】：2016
【分類號】：TP242
，

本文編號：1390948