Google开源激光SLAM算法论文原文

Real-Time Loop Closure in2D LIDAR SLAM Wolfgang Hess1,Damon Kohler1,Holger Rapp1,Daniel Andor1

Abstract—Portable laser range-?nders,further referred to as LIDAR,and simultaneous localization and mapping(SLAM) are an ef?cient method of acquiring as-built?oor plans. Generating and visualizing?oor plans in real-time helps the operator assess the quality and coverage of capture data.Build-ing a portable capture platform necessitates operating under limited computational resources.We present the approach used in our backpack mapping platform which achieves real-time mapping and loop closure at a5cm resolution.To achieve real-time loop closure,we use a branch-and-bound approach for computing scan-to-submap matches as constraints.We provide experimental results and comparisons to other well known approaches which show that,in terms of quality,our approach is competitive with established techniques.

I.I NTRODUCTION

As-built?oor plans are useful for a variety of applications. Manual surveys to collect this data for building management tasks typically combine computed-aided design(CAD)with laser tape measures.These methods are slow and,by em-ploying human preconceptions of buildings as collections of straight lines,do not always accurately describe the true nature of the https://www.wendangku.net/doc/557023132.html,ing SLAM,it is possible to swiftly and accurately survey buildings of sizes and complexities that would take orders of magnitude longer to survey manually. Applying SLAM in this?eld is not a new idea and is not the focus of this paper.Instead,the contribution of this paper is a novel method for reducing the computational requirements of computing loop closure constraints from laser range data.This technique has enabled us to map very large?oors,tens-of-thousands of square meters,while providing the operator fully optimized results in real-time.

II.R ELATED WORK

Scan-to-scan matching is frequently used to compute relative pose changes in laser-based SLAM approaches,for example[1]–[4].On its own,however,scan-to-scan matching quickly accumulates error.

Scan-to-map matching helps limit this accumulation of error.One such approach,which uses Gauss-Newton to?nd local optima on a linearly interpolated map,is[5].In the presence of good initial estimates for the pose,provided in this case by using a suf?ciently high data rate LIDAR,locally optimized scan-to-map matching is ef?cient and robust. On unstable platforms,the laser fan is projected onto the horizontal plane using an inertial measurement unit(IMU) to estimate the orientation of gravity.

Pixel-accurate scan matching approaches,such as[1], further reduce local error accumulation.Although compu-tationally more expensive,this approach is also useful for 1All authors are at Google.loop closure detection.Some methods focus on improving on the computational cost by matching on extracted features from the laser scans[4].Other approaches for loop closure detection include histogram-based matching[6],feature de-tection in scan data,and using machine learning[7].

Two common approaches for addressing the remaining local error accumulation are particle?lter and graph-based SLAM[2],[8].

Particle?lters must maintain a representation of the full system state in each particle.For grid-based SLAM,this quickly becomes resource intensive as maps become large;

e.g.one of our test cases is22,000m2collected over a3km trajectory.Smaller dimensional feature representations,such as[9],which do not require a grid map for each particle,may be used to reduce resource requirements.When an up-to-date grid map is required,[10]suggests computing submaps, which are updated only when necessary,such that the?nal map is the rasterization of all submaps.

Graph-based approaches work over a collection of nodes representing poses and features.Edges in the graph are con-straints generated from observations.Various optimization methods may be used to minimize the error introduced by all constraints,e.g.[11],[12].Such a system for outdoor SLAM that uses a graph-based approach,local scan-to-scan matching,and matching of overlapping local maps based on histograms of submap features is described in[13].

III.S YSTEM OVERVIEW

Google’s Cartographer provides a real-time solution for indoor mapping in the form of a sensor equipped backpack that generates2D grid maps with a r=5cm resolution.The operator of the system can see the map being created while walking through a https://www.wendangku.net/doc/557023132.html,ser scans are inserted into a submap at the best estimated position,which is assumed to be suf?ciently accurate for short periods of time.Scan matching happens against a recent submap,so it only depends on recent scans,and the error of pose estimates in the world frame accumulates.

To achieve good performance with modest hardware re-quirements,our SLAM approach does not employ a particle ?lter.To cope with the accumulation of error,we regularly run a pose optimization.When a submap is?nished,that is no new scans will be inserted into it anymore,it takes part in scan matching for loop closure.All?nished submaps and scans are automatically considered for loop closure.If they are close enough based on current pose estimates,a scan matcher tries to?nd the scan in the submap.If a suf?ciently good match is found in a search window around the currently estimated pose,it is added as a loop closing constraint to the

the optimization every of an operator is that loops are location is revisited.This leads that the loop closure scan than new scans are added,We achieve this by using and several precomputed grids per ?nished submap.

IV.L OCAL 2D SLAM

Our system combines separate local and global approaches to 2D SLAM.Both approaches optimize the pose,ξ=(ξx ,ξy ,ξθ)consisting of a (x,y )translation and a rotation ξθ,of LIDAR observations,which are further referred to as scans.On an unstable platform,such as our backpack,an IMU is used to estimate the orientation of gravity for projecting scans from the horizontally mounted LIDAR into the 2D world.

In our local approach,each consecutive scan is matched against a small chunk of the world,called a submap M ,using a non-linear optimization that aligns the scan with the submap;this process is further referred to as scan matching.Scan matching accumulates error over time that is later removed by our global approach,which is described in Section V.A.Scans

Submap construction is the iterative process of repeatedly aligning scan and submap coordinate frames,further referred to as frames.With the origin of the scan at 0∈R 2,we now write the information about the scan points as H ={h k }k =1,...,K ,h k ∈R 2.The pose ξof the scan frame in the submap frame is represented as the transformation T ξ,which rigidly transforms scan points from the scan frame into the submap frame,de?ned as

T ξp =

cos ξθ?sin ξθ

sin ξθcos ξθ

R ξ

p + ξx ξy t ξ

.(1)B.Submaps

A few consecutive scans are used to build a submap.These

submaps take the form of probability grids M :r Z ×r Z →[p min ,p max ]which map from discrete grid points at a given

Fig.1.Grid points and associated pixels.

resolution r ,for example 5cm,to values.These values can be

thought of as the probability that a grid point is obstructed.For each grid point,we de?ne the corresponding pixel to consist of all points that are closest to that grid point.

Whenever a scan is to be inserted into the probability grid,a set of grid points for hits and a disjoint set for misses are computed.For every hit,we insert the closest grid point into the hit set.For every miss,we insert the grid point associated with each pixel that intersects one of the rays between the scan origin and each scan point,excluding grid points which are already in the hit set.Every formerly unobserved grid point is assigned a probability p hit or p miss if it is in one of these sets.If the grid point x has already been observed,we update the odds for hits and misses as

odds (p )=

p 1?p

,(2)M new (x )=clamp(odds ?1(odds(M old (x ))·odds (p hit )))

(3)and equivalently for misses.

Fig.2.A scan and pixels associated with hits (shaded and crossed out)and misses (shaded only).

C.Ceres scan matching

Prior to inserting a scan into a submap,the scan pose ξis optimized relative to the current local submap using a Ceres -based [14]scan matcher.The scan matcher is responsible for ?nding a scan pose that maximizes the probabilities at the scan points in the submap.We cast this as a nonlinear least squares problem

argmin

ξ

K k =1

1?M smooth (T ξh k )

2

(CS)

where T ξtransforms h k from the scan frame to the submap frame according to the scan pose.The function M smooth :R 2→R is a smooth version of the probability values in the local submap.We use bicubic interpolation.As a result,values outside the interval [0,1]can occur but are considered harmless.

Mathematical optimization of this smooth function usually gives better precision than the resolution of the grid.Since this is a local optimization,good initial estimates are re-quired.An IMU capable of measuring angular velocities can be used to estimate the rotational component θof the pose

between scan matches.A higher frequency of scan matches or a pixel-accurate scan matching approach,although more computationally intensive,can be used in the absence of an IMU.

V.C LOSING LOOPS

As scans are only matched against a submap containing a few recent scans,the approach described above slowly accumulates error.For only a few dozen consecutive scans,the accumulated error is small.

Larger spaces are handled by creating many small sub-maps.Our approach,optimizing the poses of all scans and submaps,follows Sparse Pose Adjustment [2].The relative poses where scans are inserted are stored in memory for use in the loop closing optimization.In addition to these relative poses,all other pairs consisting of a scan and a submap are considered for loop closing once the submap no longer changes.A scan matcher is run in the background and if a good match is found,the corresponding relative pose is added to the optimization problem.A.Optimization problem

Loop closure optimization,like scan matching,is also formulated as a nonlinear least squares problem which allows easily adding residuals to take additional data into account.Once every few seconds,we use Ceres [14]to compute a solution to

argmin

Ξm ,Ξs

12 ij ρ E 2(ξm i ,ξs

j ;Σij ,ξij ) (SPA)where the submap poses Ξm ={ξm

i }i =1,...,m and the scan

poses Ξs ={ξs

j }j =1,...,n in the world are optimized given some constraints .These constraints take the form of relative poses ξij and associated covariance matrices Σij .For a pair of submap i and scan j ,the pose ξij describes where in the submap coordinate frame the scan was matched.The covariance matrices Σij can be evaluated,for example,fol-lowing the approach in [15],or locally using the covariance estimation feature of Ceres [14]with (CS).The residual E for such a constraint is computed by

E 2(ξm i ,ξs j ;Σij ,ξij )=e (ξm i ,ξs j ;ξij )T Σ?1ij e (ξm i ,ξs

j ;ξij ),(4)e (ξm i ,ξs j ;ξij )=ξij ? R ?1ξm i (t ξm i ?t ξs j )ξm i ;θ?ξs j ;θ

.(5)A loss function ρ,for example Huber loss ,is used to reduce the in?uence of outliers which can appear in (SPA)when scan matching adds incorrect constraints to the opti-mization problem.For example,this may happen in locally symmetric environments,such as of?ce cubicles.Alternative approaches to outliers include [16].B.Branch-and-bound scan matching

We are interested in the optimal,pixel-accurate match

ξ

=argmax

ξ∈W

K k =1

M nearest (T ξh k ),

(BBS)

where W is the search window and M nearest is M extended

to all of R 2by rounding its arguments to the nearest grid point ?rst,that is extending the value of a grid points to the corresponding pixel.The quality of the match can be improved further using (CS).

Ef?ciency is improved by carefully choosing step sizes.We choose the angular step size δθso that scan points at the maximum range d max do not move more than r ,the width of one https://www.wendangku.net/doc/557023132.html,ing the law of cosines,we derive

d max =

max k =1,...,K

h k ,(6)δθ=arccos(1?r 2

2d 2max

).

(7)

We compute an integral number of steps covering given linear and angular search window sizes,e.g.,W x =W y =7m and W θ=30?,

w x = W x r ,w y = W y r ,w θ= W θ

δθ .(8)

This leads to a ?nite set W forming a search window

around an estimate ξ0placed in its center,

W ={?w x ,...,w x }×{?w y ,...,w y }×{?w θ,...,w θ},

(9)W ={ξ0+(rj x ,rj y ,δθj θ):(j x ,j y ,j θ)∈W}.

(10)

A naive algorithm to ?nd ξ can easily be formulated,see Algorithm 1,but for the search window sizes we have in mind it would be far too slow.

Algorithm 1Naive algorithm for (BBS)best score ←?∞

for j x =?w x to w x do for j y =?w y to w y do for j θ=?w θto w θdo score ← K

k =1M nearest (T ξ0+(rj x ,rj y ,δθj θ)h k )if score >best score then

match ←ξ0+(rj x ,rj y ,δθj θ)best score ←score end if end for end for end for

return best score and match when set.Instead,we use a branch and bound approach to ef?ciently compute ξ over larger search windows.See Algorithm 2for the generic approach.This approach was ?rst suggested in the context of mixed integer linear programs [17].Literature on the topic is extensive;see [18]for a short overview.The main idea is to represent subsets of possibilities as nodes in a tree where the root node represents all possible solutions,W in our case.The children of each node form a partition of their parent,so that they together represent the same set of possibilities.The leaf nodes are singletons;each represents a single feasible solution.Note that the algorithm is exact.It provides the same solution as the naive approach,

as long as the score(c)of inner nodes c is an upper bound on the score of its elements.In that case,whenever a node is bounded,a solution better than the best known solution so far does not exist in this subtree.

To arrive at a concrete algorithm,we have to decide on the method of node selection,branching,and computation of upper bounds.

1)Node selection:Our algorithm uses depth-?rst search (DFS)as the default choice in the absence of a better alternative:The ef?ciency of the algorithm depends on a large part of the tree being pruned.This depends on two things:a good upper bound,and a good current solution.The latter part is helped by DFS,which quickly evaluates many leaf nodes.Since we do not want to add poor matches as loop closing constraints,we also introduce a score threshold below which we are not interested in the optimal solution. Since in practice the threshold will not often be surpassed, this reduces the importance of the node selection or?nding an initial heuristic solution.Regarding the order in which the children are visited during the DFS,we compute the upper bound on the score for each child,visiting the most promising child node with the largest bound?rst.This method is Algorithm3.

2)Branching rule:Each node in the tree is described by

a tuple of integers c=(c x,c y,cθ,c h)∈Z4.Nodes at height c h combine up to2c h×2c h possible translations but represent a speci?c rotation:

W c=

(j x,j y)∈Z2:(11)

c x≤j x

c y≤j y

×{cθ}

,

W c=W c∩W.(12)

Algorithm2Generic branch and bound

best score←?∞

C←C0

while C=?do

Select a node c∈C and remove it from the set.

if c is a leaf node then

if score(c)>best score then

solution←n

best score←score(c)

end if

else

if score(c)>best score then

Branch:Split c into nodes C c.

C←C∪C c

else

Bound.

end if

end if

end while

return best score and solution when set.Algorithm3DFS branch and bound scan matcher for(BBS) score threshold

Compute and memorize a score for each element in C0. Initialize a stack C with C0sorted by score,the maximum score at the top.

while C is not empty do

Pop c from the stack C.

if score(c)>best score then

if c is a leaf node then

match←ξc

best score←score(c)

else

Branch:Split c into nodes C c.

Compute and memorize a score for each element

in C c.

Push C c onto the stack C,sorted by score,the

maximum score last.

end if

end if

end while

return best score and match when set.

Leaf nodes have height c h=0,and correspond to feasible solutions W ξc=ξ0+(rc x,rc y,δθcθ).

In our formulation of Algorithm3,the root node,encom-passing all feasible solutions,does not explicitly appear and branches into a set of initial nodes C0at a?xed height h0 covering the search window

W0,x={?w x+2h0j x:j x∈Z,0≤2h0j x≤2w x},

W0,y={?w y+2h0j y:j y∈Z,0≤2h0j y≤2w y},

W0,θ={jθ∈Z:?wθ≤jθ≤wθ},

C0=W0,x×W0,y×W0,θ×{h0}.

(13)

At a given node c with c h>1,we branch into up to four children of height c h?1

C c=

{c x,c x+2c h?1}×{c y,c y+2c h?1}

×cθ

∩W

×{c h?1}.

(14)

3)Computing upper bounds:The remaining part of the branch and bound approach is an ef?cient way to compute upper bounds at inner nodes,both in terms of computational effort and in the quality of the bound.We use

score(c)=

K

k=1

max

j∈W c

M nearest(Tξ

j

h k)

≥

K

k=1

max

j∈W c

M nearest(Tξ

j

h k)

≥max

j∈W c

K

k=1

M nearest(Tξ

j

h k).

(15)

To be able to compute the maximum ef?ciently,we use precomputed grids M c h precomp.Precomputing one grid per possible height c h allows us to compute the score with effort

linear in the number of scan points.Note that,to be able to do this,we also compute the maximum over W c which can be larger than W c near the boundary of our search space.

score (c )=

K k =1

M c h

precomp (T ξc h k ),

(16)M h

precomp (x,y )

=

max

x ∈[x,x +r (2h ?1)]y ∈[y,y +r (2h ?1)]

M nearest (x ,y )

(17)

with ξc as before for the leaf nodes.Note that M h

precomp has the same pixel structure as M nearest ,but in each pixel storing the maximum of the values of the 2h ×2h box

of pixels

beginning

there.An example of such precomputed grids is given in Figure 3.

Fig.3.Precomputed grids of size 1,4,16and 64.

To keep the computational effort for constructing the precomputed grids low,we wait until a probability grid will receive no further updates.Then we compute a collection of precomputed grids,and start matching against it.

For each precomputed grid,we compute the maximum of a 2h pixel wide row starting at each https://www.wendangku.net/doc/557023132.html,ing this inter-mediate result,the next precomputed grid is then constructed.The maximum of a changing collection of values can be kept up-to-date in amortized O (1)if values are removed in the order in which they have been added.Successive maxima are kept in a deque that can be de?ned recursively as containing the maximum of all values currently in the collection followed by the list of successive maxima of all values after the ?rst occurrence of the maximum.For an empty collection of values,this list is https://www.wendangku.net/doc/557023132.html,ing this approach,the precomputed grids can be computed in O (n )where n is the number of pixels in each precomputed grids.An alternative way to compute upper bounds is to compute lower resolution probability grids,successively halving the resolution,see [1].Since the additional memory consumption of our approach is acceptable,we prefer it over using lower resolution probability grids which lead to worse bounds than (15)and thus negatively impact performance.

VI.E XPERIMENTAL RESULTS

In this section,we present some results of our SLAM al-gorithm computed from recorded sensor data using the same online algorithms that are

used interactively on the backpack.First,we show results using data collected by the sensors

Fig.4.Cartographer map of the 2nd ?oor of the Deutsches Museum.

of our Cartographer backpack in the Deutsches Museum in Munich.Second,we demonstrate that our algorithms work well with inexpensive hardware by using data collected from a robotic vacuum cleaner https://www.wendangku.net/doc/557023132.html,stly,we show results using the Radish data set [19]and compare ourselves to published results.

A.Real-World Experiment:Deutsches Museum

Using data collected at the Deutsches Museum spanning 1,913s of sensor data or 2,253m (according to the computed solution),we computed the map shown in Figure 4.On a workstation with an Intel Xeon E5-1650at 3.2GHz,our SLAM algorithm uses 1,018s CPU time,using up to 2.2GB of memory and up to 4background threads for loop closure scan matching.It ?nishes after 360s wall clock time,meaning it achieved 5.3times real-time performance.

The generated graph for the loop closure optimization consists of 11,456nodes and 35,300edges.The optimization problem (SPA)is run every time a few nodes have been added to the graph.A typical solution takes about 3itera-tions,and ?nishes in about 0.3s.

Fig.5.Cartographer map generated using Revo LDS sensor data.

TABLE I

Q UANTITATIVE ERRORS WITH R EVO LDS

Laser Tape Cartographer Error(absolute)Error(relative)

4.094.08?0.01?0.2%

5.405.43+0.03+0.6%

8.678.74+0.07+0.8%

15.0915.20+0.11+0.7%

15.1215.23+0.11+0.7%

B.Real-World Experiment:Neato’s Revo LDS

Neato Robotics uses a laser distance sensor(LDS)called Revo LDS[20]in their vacuum cleaners which costs under $30.We captured data by pushing around the vacuum cleaner on a trolley while taking scans at approximately2Hz over its debug connection.Figure5shows the resulting5cm resolution?oor plan.To evaluate the quality of the?oor plan, we compare laser tape measurements for5straight lines to the pixel distance in the resulting map as computed by a drawing tool.The results are presented in Table I,all values are in meters.The values are roughly in the expected order of magnitude of one pixel at each end of the line.

https://www.wendangku.net/doc/557023132.html,parisons using the Radish data set

We compare our approach to others using the benchmark measure suggested in[21],which compares the error in rela-tive pose changes to manually curated ground truth relations. Table II shows the results computed by our Cartographer SLAM algorithm.For comparison,we quote results for Graph Mapping(GM)from[21].Additionally,we quote more recently published results from[9]in Table III.All errors are given in meters and degrees,either absolute or squared,together with their standard deviation.

Each public data set was collected with a unique sensor con?guration that differs from our Cartographer backpack. Therefore,various algorithmic parameters needed to be adapted to produce reasonable results.In our experience,tun-ing Cartographer is only required to match the algorithm to the sensor con?guration and not to the speci?c surroundings.

TABLE II

Q UANTITATIVE COMPARISON OF ERROR WITH[21]

Cartographer GM Aces

Absolute translational0.0375±0.04260.044±0.044

Squared translational0.0032±0.02850.004±0.009

Absolute rotational0.373±0.4690.4±0.4

Squared rotational0.359±3.6960.3±0.8

Intel

Absolute translational0.0229±0.02390.031±0.026

Squared translational0.0011±0.00400.002±0.004

Absolute rotational0.453±1.3351.3±4.7

Squared rotational1.986±23.98824.0±166.1

MIT Killian Court

Absolute translational0.0395±0.04880.050±0.056

Squared translational0.0039±0.01440.006±0.029

Absolute rotational0.352±0.3530.5±0.5

Squared rotational0.248±0.6100.9±0.9

MIT CSAIL

Absolute translational0.0319±0.03630.004±0.009

Squared translational0.0023±0.00990.0001±0.0005

Absolute rotational0.369±0.3650.05±0.08

Squared rotational0.270±0.6370.01±0.04

Freiburg bldg79

Absolute translational0.0452±0.03540.056±0.042

Squared translational0.0033±0.00550.005±0.011

Absolute rotational0.538±0.7180.6±0.6

Squared rotational0.804±3.6270.7±1.7

Freiburg hospital(local)

Absolute translational0.1078±0.19430.143±0.180

Squared translational0.0494±0.28310.053±0.272

Absolute rotational0.747±2.0470.9±2.2

Squared rotational4.745±40.0815.5±46.2

Freiburg hospital(global)

Absolute translational5.2242±6.623011.6±11.9

Squared translational71.0288±267.7715276.1±516.5

Absolute rotational3.341±4.7976.3±6.2

Squared rotational34.107±127.22777.2±154.8 Since each public data set has a unique sensor con?g-uration,we cannot be sure that we did not also?t our parameters to the speci?c locations.The only exception being the Freiburg hospital data set where there are two separate relations?les.We tuned our parameters using the local relations but also see good results on the global relations.

TABLE III

Q UANTITATIVE COMPARISON OF ERROR WITH[9]

Cartographer Graph FLIRT Intel

Absolute translational0.0229±0.02390.02±0.02

Absolute rotational0.453±1.3350.3±0.3

Freiburg bldg79

Absolute translational0.0452±0.03540.06±0.09

Absolute rotational0.538±0.7180.8±1.1

Freiburg hospital(local)

Absolute translational0.1078±0.19430.18±0.27

Absolute rotational0.747±2.0470.9±2.0

Freiburg hospital(global)

Absolute translational5.2242±6.62308.3±8.6

Absolute rotational3.341±4.7975.0±5.3

TABLE IV

L OOP CLOSURE PRECISION

Test case No.of constraints Precision

Aces97198.1%

Intel578697.2%

MIT Killian Court91693.4%

MIT CSAIL185794.1%

Freiburg bldg7941299.8%

Freiburg hospital55477.3%

TABLE V

P ERFORMANCE

Test case Data duration(s)Wall clock(s)

Aces136641

Intel2691179

MIT Killian Court7678190

MIT CSAIL42435

Freiburg bldg79106162

Freiburg hospital482010

The most signi?cant differences between all data sets is the frequency and quality of the laser scans as well as the availability and quality of odometry.

Despite the relatively outdated sensor hardware used in the public data sets,Cartographer SLAM consistently performs within our expectations,even in the case of MIT CSAIL, where we perform considerably worse than Graph Mapping. For the Intel data set,we outperform Graph Mapping,but not Graph FLIRT.For MIT Killian Court we outperform Graph Mapping in all metrics.In all other cases,Cartographer outperforms both Graph Mapping and Graph FLIRT in most but not all metrics.

Since we add loop closure constraints between submaps and scans,the data sets contain no ground truth for them.It is also dif?cult to compare numbers with other approaches based on scan-to-scan.Table IV shows the number of loop closure constraints added for each test case(true and false positives),as well as the precision,that is the fraction of true positives.We determine the set of true positive constraints to be the subset of all loop closure constraints which are not violated by more than20cm or1?when we compute (SPA).We see that while our scan-to-submap matching procedure produces false positives which have to be handled in the optimization(SPA),it manages to provide a suf?cient number of loop closure constraints in all test cases.Our use of the Huber loss in(SPA)is one of the factors that renders loop closure robust to outliers.In the Freiburg hospital case, the choice of a low resolution and a low minimum score for the loop closure detection produces a comparatively high rate of false positives.The precision can be improved by raising the minimum score for loop closure detection,but this decreases the solution quality in some dimensions according to ground truth.The authors believe that the ground truth remains the better benchmark of?nal map quality.

The parameters of Cartographer’s SLAM were not tuned for CPU performance.We still provide the wall clock times in Table V which were again measured on a workstation with an Intel Xeon E5-1650at3.2GHz.We provide the duration of the sensor data for comparison.

VII.C ONCLUSIONS

In this paper,we presented and experimentally validated a2D SLAM system that combines scan-to-submap match-ing with loop closure detection and graph optimization. Individual submap trajectories are created using our local, grid-based SLAM approach.In the background,all scans are matched to nearby submaps using pixel-accurate scan matching to create loop closure constraints.The constraint graph of submap and scan poses is periodically optimized in the background.The operator is presented with an up-to-date preview of the?nal map as a GPU-accelerated combination of?nished submaps and the current submap. We demonstrated that it is possible to run our algorithms on modest hardware in real-time.

A CKNOWLEDGMENTS

This research has been validated through experiments in the Deutsches Museum,Munich.The authors thank its administration for supporting our work.

Comparisons were done using manually veri?ed relations and results from[21]which uses data from the Robotics Data Set Repository(Radish)[19].Thanks go to Patrick Beeson,Dieter Fox,Dirk H¨a hnel,Mike Bosse,John Leonard, Cyrill Stachniss for providing this data.The data for the Freiburg University Hospital was provided by Bastian Steder, Rainer K¨u mmerle,Christian Dornhege,Michael Ruhnke, Cyrill Stachniss,Giorgio Grisetti,and Alexander Kleiner.

R EFERENCES

[1] E.Olson,“M3RSM:Many-to-many multi-resolution scan matching,”

in Proceedings of the IEEE International Conference on Robotics and Automation(ICRA),June2015.

[2]K.Konolige,G.Grisetti,R.K¨u mmerle,W.Burgard,B.Limketkai,

and R.Vincent,“Sparse pose adjustment for2D mapping,”in IROS, Taipei,Taiwan,10/20102010.

[3] F.Lu and https://www.wendangku.net/doc/557023132.html,ios,“Globally consistent range scan alignment for

environment mapping,”Autonomous robots,vol.4,no.4,pp.333–349,1997.

[4] F.Mart′?n,R.Triebel,L.Moreno,and R.Siegwart,“Two different

tools for three-dimensional mapping:DE-based scan matching and feature-based loop detection,”Robotica,vol.32,no.01,pp.19–41, 2014.

[5]S.Kohlbrecher,J.Meyer,O.von Stryk,and U.Klingauf,“A?exible

and scalable SLAM system with full3D motion estimation,”in Proc.IEEE International Symposium on Safety,Security and Rescue Robotics(SSRR).IEEE,November2011.

[6]M.Himstedt,J.Frost,S.Hellbach,H.-J.B¨o hme,and E.Maehle,

“Large scale place recognition in2D LIDAR scans using geometrical landmark relations,”in Intelligent Robots and Systems(IROS2014), 2014IEEE/RSJ International Conference on.IEEE,2014,pp.5030–5035.

[7]K.Granstr¨o m,T.B.Sch¨o n,J.I.Nieto,and F.T.Ramos,“Learning to

close loops from range data,”The International Journal of Robotics Research,vol.30,no.14,pp.1728–1754,2011.

[8]G.Grisetti, C.Stachniss,and W.Burgard,“Improving grid-based

SLAM with Rao-Blackwellized particle?lters by adaptive proposals and selective resampling,”in Robotics and Automation,2005.ICRA 2005.Proceedings of the2005IEEE International Conference on.

IEEE,2005,pp.2432–2437.

[9]G.D.Tipaldi,M.Braun,and K.O.Arras,“FLIRT:Interest regions for

2D range data with applications to robot navigation,”in Experimental Robotics.Springer,2014,pp.695–710.

[10]J.Strom and E.Olson,“Occupancy grid rasterization in large environ-

ments for teams of robots,”in Intelligent Robots and Systems(IROS), 2011IEEE/RSJ International Conference on.IEEE,2011,pp.4271–4276.

[11]R.K¨u mmerle,G.Grisetti,H.Strasdat,K.Konolige,and W.Burgard,

“g2o:A general framework for graph optimization,”in Robotics and Automation(ICRA),2011IEEE International Conference on.IEEE, 2011,pp.3607–3613.

[12]L.Carlone,R.Aragues,J. A.Castellanos,and B.Bona,“A fast

and accurate approximation for planar pose graph optimization,”The International Journal of Robotics Research,pp.965–987,2014. [13]M.Bosse and R.Zlot,“Map matching and data association for large-

scale two-dimensional laser scan-based SLAM,”The International Journal of Robotics Research,vol.27,no.6,pp.667–691,2008. [14]S.Agarwal,K.Mierle,and Others,“Ceres solver,”http://ceres-solver.

org.

[15] E. B.Olson,“Real-time correlative scan matching,”in Robotics

and Automation,2009.ICRA’09.IEEE International Conference on.

IEEE,2009,pp.4387–4393.[16]P.Agarwal,G.D.Tipaldi,L.Spinello,C.Stachniss,and W.Burgard,

“Robust map optimization using dynamic covariance scaling,”in Robotics and Automation(ICRA),2013IEEE International Conference on.IEEE,2013,pp.62–69.

[17] https://www.wendangku.net/doc/557023132.html,nd and A.G.Doig,“An automatic method of solving discrete

programming problems,”Econometrica,vol.28,no.3,pp.497–520, 1960.

[18]J.Clausen,“Branch and bound algorithms-principles and examples,”

Department of Computer Science,University of Copenhagen,pp.1–30,1999.

[19] A.Howard and N.Roy,“The robotics data set repository(Radish),”

2003.[Online].Available:https://www.wendangku.net/doc/557023132.html,/

[20]K.Konolige,J.Augenbraun,N.Donaldson,C.Fiebig,and P.Shah,

“A low-cost laser distance sensor,”in Robotics and Automation,2008.

ICRA2008.IEEE International Conference on.IEEE,2008,pp.

3002–3008.

[21]R.K¨u mmerle, B.Steder, C.Dornhege,M.Ruhnke,G.Grisetti,

C.Stachniss,and A.Kleiner,“On measuring the accuracy of SLAM

algorithms,”Autonomous Robots,vol.27,no.4,pp.387–407,2009.

用google查找文献资料方法-整理版

1.只需要在搜索里面加上inurl:pdf然后后面加上杂志的限制，比如"by cell press"，就可以查到这个期刊的全部论文了。 2.基本搜索：+，-，OR 2.1.表示逻辑“+”：只要空格就可以了。 2.2.表示逻辑“非”操作：用减号“-”。 注意：这里的“+”和“-”号，是英文字符，操作符与作用的关键字之间，不能有空格。比如避免“易筋经- 吸星大法”。 2.3.GOOGLE用大写的“OR”表示逻辑“或”操作。 示例：搜索包含布兰妮“Britney”或者披头士“Beatles”、或者两者均有的中文网页。 搜索：“britney OR beatles”，结果：已搜索有关britney OR beatles的中文(简体)网页。共约有14,600项查询结果，这是第1-10项。搜索用时0.08秒。搜索：“布兰妮OR 披头士”结果：找不到和您的查询-布兰妮OR 披头士-相符的网页。 注意：小写的“or”，在查询的时候将被忽略；这样上述的操作实际上变成了一次“与”查询 3.辅助搜索：通配符、大小写、句子、忽略字符以及强制搜索 GOOGLE不支持通配符，如“*”、“?”等，只能做精确查询，关键字后面的“*”或者“?”会被忽略掉。GOOGLE对英文字符大小写不敏感，“GOD”和“god”搜索的结果是一样的。GOOGLE的关键字可以是词组（中间没有空格），也可以是句子（中间有空格），但是，用句子做关键字，必须加英文引号。 4.高级搜索：site，link，inurl，allinurl，intitle，allintitle 4.1.“site” 表示搜索结果局限于某个具体网站或者网站频道，如“https://www.wendangku.net/doc/557023132.html,”、“https://www.wendangku.net/doc/557023132.html,”，或者是某个域名，如“https://www.wendangku.net/doc/557023132.html,”、“com”等等。如果是要排除某网站或者域名范围内的页面，只需用“-网站/域名”。示例：搜索中文教育科研网站（https://www.wendangku.net/doc/557023132.html,）上所有包含“金庸”的页面。搜索：“金庸site:https://www.wendangku.net/doc/557023132.html,”结果：已搜索有关金庸site:https://www.wendangku.net/doc/557023132.html,的中文(简体)网页。共约有2,680项查询结果，这是第1-10项。搜索用时0.31秒。 示例：搜索包含“金庸”和“古龙”的中文新浪网站页面，搜索：“金庸古龙site:https://www.wendangku.net/doc/557023132.html,”结果：已在https://www.wendangku.net/doc/557023132.html,搜索有关金庸古龙的中文(简体)网页。共约有869项查询结果，这是第1-10项。搜索用时0.34秒。注意：site后的冒号为英文字符，而且，冒号后不能有空格，否

利用Google进行专题信息检索的方法和技巧

利用Google进行专题信息检索的方 法和技巧 摘要随着科学研究所依赖的各种信息资源的大规模网络化数字化，搜索引擎逐渐成为网络时代的最快捷方便的个性化信息服务系统。Google成为目前最受欢迎的搜索引擎，本文全面详细总结了利用Google进行专题信息检索的方法和技巧。 关键词个性化信息服务信息检索专题检索搜索Google 方法技巧 目前，科学研究依赖的各种信息资源，包括文摘索引、期刊论文、预印本、技术报告、学位论文、会议论文、以及部分重要工具书和专著等在内的主流科研信息资源已经逐步数字化，开始形成一个逐步完善的数字化信息资源空间，科研人员可以通过网络跨时空的进行专题信息检索，获取相关科研文献资源[1]。 基于网页内容的全文检索技术是搜索引擎的核心，搜索引擎也是全自动的软件服务。从目前来看，搜索引擎，尤其是Google已成为网络时代最快捷方便的个性化信息服务系统和服务方式。[2] 1 Google成为目前最受欢迎的搜索引擎 Google(https://www.wendangku.net/doc/557023132.html,)是当今一个优秀的搜索引擎，其功能强大、特点突出、技术先进和服务优良，它在业界评测中获得多项大奖，各大引擎竞相模仿其功能和特色。Google非中国本土公司，但它支持中文搜索，其中文搜索引擎是收集亚洲网站最多的搜索引擎之一，国内使用其独立搜索引擎的人数急剧增长。 目前，全世界访问量最大的4个网站中，3家采用了Google的搜索技术，80%的互联网搜索是通过Google或使用Google技术的网站完成的。目前Google每个月接待来自世界各地的超过2800万独立访问者，全球网民通过Google可以使用86种语言，搜索30多亿个网页及其网页快照，以及4亿多张图片，每个月Google被用户使用的时间为1500万小时左右。据搜索引擎观察者网络杂志统计结果显示，至2002年10月份，网民使用Google的时间量每月达到1610万小时；相比之下，雅虎只吸引了660万小时，微软MSN仅有520万小时[3]。 2 利用Google进行专题信息检索的方法和技巧 关键词检索功能是网络信息检索工具的基本检索功能，也是Google最基本的检索功能。关键词属于自然语言，灵活、不受词表控制，但简单的关键词检索方法，命中过多，查准率很低，Google为改善关键词检索性能，提供了按相关度排列结果、布尔逻辑检索，短语或者句子检索、加权检索和限制检索等增强措施。 利用Google进行专题信息检索，为提高查准率，须认真分析课题，选择恰当的关键词，掌握和运用Google检索语法规则，准确设计表达需求的检索式，反复调整检索策略，才能获得高质量的检索结果。 2.1 简单专题信息检索，最直截了当就是在搜索框内输入一个关键词，然后点击下面的

谷歌学术论文高级搜索技巧

"+" 操作符确保您的搜索结果中包括Google 学术搜索技术通常忽略的普通字词、字母或数字，如[+de knuth]; "-" 操作符排除所有包括搜索字词的结果，如[Flowers -作者:Flowers ]； 短语搜索只返回包括这一确切短语的结果，如["随你便"]； "OR" 操作符返回包括搜索字词之一的结果，如[股票看涨期权OR 看跌期权]； "标题："操作符如[在标题：mars] 得到的结果只包括文件名中的搜索字词。 6.1、搜索结果要求包含两个及两个以上关键字 一般搜索引擎需要在多个关键字之间加上“+”，而GOOGLE无需用明文的“+”来表示逻辑“与”操作，只要空格就可以了。 示例：搜索所有包含关键词“易筋经”和“吸星大法”的中文网页 搜索：“易筋经吸星大法” 结果：已搜索有关易筋经吸星大法的中文(简体)网页。共约有774项查询结果，这是第1-10项。搜索用时0.24秒。 注意：文章中搜索语法外面的引号仅起引用作用，不能带入搜索栏内。 6.2、搜索结果要求不包含某些特定信息 GOOGLE用减号“-”表示逻辑“非”操作。 示例：搜索所有包含“易筋经”而不含“吸星大法”的中文网页 搜索：“易筋经-吸星大法” 结果：已搜索有关易筋经-吸星大法的中文(简体)网页。共约有5,150项查询结果，这是第1-10项。搜索用时0.40秒。 注意：这里的“+”和“-”号，是英文字符，而不是中文字符的“＋”和“－”。此外，操作符与作用的关键字之间，不能有空格。比如“易筋经- 吸星大法”，搜索引擎将视为关键字为易筋经和吸星大法的逻辑“与”操作，中间的“-”被忽略。 6.3、搜索结果至少包含多个关键字中的任意一个 GOOGLE用大写的“OR”表示逻辑“或”操作。假定你是布兰妮和批头士的歌迷，现在要查找所有关于布兰妮和批头士的中文网页。 示例：搜索包含布兰妮“Britney”或者披头士“Beatles”、或者两者均有的中文网页。 搜索：“Britney OR Beatles OR 布兰妮OR 批头士” 结果：已搜索有关Britney OR Beatles OR 布兰妮OR 批头士的中文(简体)网页。共约有31,300项查询结果，这是第1-10项。 注意：小写的“or”，在查询的时候将被忽略；这样上述的操作实际上变成了一次“与”查询。

google检索密码的方法

google检索密码的方法2009/12/13 12:40google检索密码的方法 inurl: inurl: 是骇客重要的搜寻方法，可搜到网址包括的关键字， 例如填上 allinurl:login password 作搜寻，便会很易找到有 login 和 password 的网页。 ------------------------------------------------------------------------------------------- file无效: file无效: 是骇客专用语法，例如想找 mdb 的数据库档案，可用 password file无效:mdb 作搜寻， 便会找到密码文件，进阶用法可配合 inurl: 使用，例如 girl file无效:jpg site:com 便可搜到所有标 题 .com 网站，而档案为 girl.jp 或网页内容有 girl 字串的。 ------------------------------------------------------------------------------------------- Index of /admin 搜到的结果大多数是容野峇嵺@ index browsing 的网站，随便按下一个连结便看到网站的资料夹和 档案分布。 ------------------------------------------------------------------------------------------- "Index of /" +password.txt 有些站长会将密码储存成 password.txt 档案，如配合 index browsing 的弁遄A将 google 的关 键字串成 "Index of /" +password.txt 作搜寻，便找到很多 password.txt ------------------------------------------------------------------------------------------- 以下还有更多输入搜寻法，有时间可自行玩玩！ "Index of /admin" "Index of /password" "Index of /mail"

google搜索技巧总结

GOOGLE 搜索技巧总结 技巧一：使用正确的关键词 针对搜索所在地区对我们产品的用法，选择正确的关键词，评估一下搜索结果页面上的匹配程度。如果一开始的结果与你想要的不一致，应立刻转向更合适的关键词再进行搜索。 技巧二：合理利用“OR”的搜索 搜索包括一个或另一个词的页面，但不一定是都包括二者，输入“OR”，确保所输入的要大写，例如：gas burner OR cast iron gas burner 在www.google.is中搜索，发现cast iron gas burner相关性更高。 技巧三：你的搜索中包括或不包括的词 Google会自动地将这些在你输入的搜索要求中的不重要的、普通的词忽略掉。这些被称作是“忽略的单词”，包括“and”、“the，”、“where”、“how”、“what”、“or”（所有字母皆为小写，还有其它一些类似的词——包括一些单独的数字或单独的字母。如果你想要让这些一般的词包含在你的搜索要求内，你可以在你确实需要的词之前加一个空格符再加上一个“+”。从另一方面来说，有时你会想要通过排除一些包含特定词的页面来精炼你的搜索结果。你可以通过使用一个“-”号来去掉搜索结果中不想要包括在内的词；在你想去掉的关键词之前加一个空格符再加上“-”，这个词会自动地排除在搜索结果之外。 技巧四：搜索近似的词 搜索与你的关键词相近的词，可以找出几乎所有与我们这种产品相关的词的用法。只要在想要搜索的词之前加上“~”符号，Google就会搜索所有包括这个词以及合适的近义词的页面。例如，在www.google.is中输入 ~gas burner得到Газоваягорелка，propane burner，fuel burner等相近关键词的搜索结果。 在此还有个额外的技巧：如果要只是列出近义词的页面，而不需要给出许多原先输入的那个词的页面，可以用“-”符号来连接“~”操作，就能在近义词所得的结果中排除原先输入的词。例如“~gas burner –gas burner” 技巧五：搜索特定的词组 当你搜索一个特定词组时，如果你只是简单地输入词组中所有的词你是无法得到最好的结果的。Google也许能够反馈出包含这个词组的结果，但它也会列出包含你所输入所有词的结果，却未必让这些词按照正确的顺序。如果你要搜索一个特定的词组，你应该将整个词组放在一个引号内。这样就能让Google搜索规定顺序的精确的关键词。例如，Homeport Canada Inc 在google中要加“”才能搜索到相关信息。 技巧六：列出相似的页面 Related:这个操作算符所显示的页面会与特定的页面在某些方面是相似的。例如，related:https://www.wendangku.net/doc/557023132.html,搜索结果有ec21, globalsources, made-in-china, tradekey, tradeindia,ecplaza等。 技巧七：网页翻译 GOOGLE提供了网页翻译功能！！虽然目前只支持有限的拉丁语、法语、西班牙语、德语和葡萄牙文。 技巧八：对搜索的网站进行限制 “site:”表示搜索结果局限于某个具体网站或者网站频道或者是某个域名，如果是要排除某网站或者域名范围内的页面，只需用“-网站/域名”。例如：将gas burner –site:https://www.wendangku.net/doc/557023132.html,输入google中搜索到的结果为除去由阿里巴巴网站提供的产品。注意：site后的冒号为英文字符，而且，冒号后不能有空格，否则，“site:”将被作为一个搜索的关键字。此外，网站域名不能有“http”以及“www”前缀，也不能有任何“/”的目录后

google搜索国外客户技巧

用google搜索国外客户的三十绝招https://www.wendangku.net/doc/557023132.html, 2009年06月29日10:37 中国B2B 研究中心发表评论联系我们 1、在Google中输入产品名称+importers。(也可以用importer代替importers进行搜索。不同的产品或者行业，这些网站的排名往往不太一样，大家要是用自己的产品测试，应选取排名比较靠前的网站加以利用。) 2、关健词上加引号，即搜索“Product-A importer”，在输入时将引号一起输入。这种方法可以保障在搜索出来的网页中我们输入的关键词是连接在一起的，不像上一种方法得到的结果中那样，输入的关键词可能是分开的。这样搜索结果虽然数量上大大降低，但准确性必然大大提高。 3、搜索产品名称+distributor，搜索时如果加上引号，能得到更准确的结果。虽然这样做可能牺牲很多潜在客户，但如果运气好的话就可以找到很多分销商的信息。 4、其他类型目标客户搜索：产品名称+其他客户类型。相关目标客户的词语还包括：buyer, company, wholesaler, retailer, supplier, vendor及复数形式，可以用来和产品名称结合搜索。这样搜索的结果不会很多，但包含比较丰富的客户信息和其他市场信息，比如行业状况、竞争对手信息和技术资料等。 5、Price +产品名称。通过这种方法得到的信息，其中一部分往往能让你找到很多的在网上销售产品的零售商和经销商，还有一部分搜索结果是一些市场报告、谈论产品行情的文章。如果是比较新的资料可以作为参考。 6、搜索buy +产品名称。这种方法可以帮助你发现可能被我们忽略的求购信息。 7、国家名称限制方法。在前面6种方法的基础上加入国家名称限制。一般从这种搜索结果中我们可以得到我们关心的产品在目标市场的情况，其中也包含不少客户信息和客户信息源。 8、关联产品法。产品名称+关联产品名称。这样的搜索结果往往是一些目标客户网站和行业网站。 9、著名买家法。产品名称+你的行业里面著名买家的公司简称或者全称。这种方法可以帮助我们找到行业市场的情况，并能在相关的网站中找到其他买家的名字。 10、Market research方法。产品名称+ Market research。这种方法用以搜索某种产品的市场研究报告。一般在这种报告的提要或者内容中，可能会提到很多著名的行业内的公司，包括制造商和分销商。 11、观察搜索引擎右侧广告。搜索产品名称后，查看搜索结果右侧广告。我们经常可以看到在Google右侧会出现一些文字广告。这是Google为了防止影响搜索结果的公正性而特别置于右侧的，这种方式既照顾到了搜索人不想受广告干扰的心理，也照顾到了广告主的利益。当我们根据以上很多的关键词搜索目标客户信息时，往往那些广告主提供的服务也是值得我们关注的。

google检索方法

利用Google进行专题信息检索的方法和技巧 关键词检索功能是网络信息检索工具的基本检索功能，也是Google最基本的检索功能。关键词属于自然语言，灵活、不受词表控制，但简单的关键词检索方法，命中过多，查准率很低，Google为改善关键词检索性能，提供了按相关度排列结果、布尔逻辑检索，短语或者句子检索、加权检索和限制检索等增强措施。 利用Google进行专题信息检索，为提高查准率，须认真分析课题，选择恰当的关键词，掌握和运用Google检索语法规则，准确设计表达需求的检索式，反复调整检索策略，才能获得高质量的检索结果。 2.1 简单专题信息检索，最直截了当就是在搜索框内输入一个关键词，然后点击下面的“Google搜索”按钮（或者直接回车），结果就出来了。 如果检索人员或用户对查询的领域熟悉，只想寻找某些专题网站，首先考虑用目录检索，Google根据其专业的“网页级别”（PageRank）技术对目录中登录的网站进行了排序，可以使检索具更高效率，按所需主题确定沿某类层层查找网站，目录分类明确，网站专题信息集中，剔除了大量不相关的信息，不过对查找中文信息，Google的中文目录太少，只有非常普通简单的类目，可能很难满足要求。 2.2 熟练掌握Google的基本搜索：+，-，OR ，学会使用两个关键字进行复杂专题信息检索

检索复杂专题依靠单个关键词查准率很低，要提高查全检准率，需进行详细的主题分析，选择多个关键词构造检索式。要分清主要概念和次要概念，去掉被隐含了的概念，确定需要排除的某些概念和不宜选用的泛指概念，以便在制定检索策略时有所侧重，保证检索提问的确切表达。正确选择关键词，各种类型的检索课题对检索的查全率和查准率有着不同的要求，可以增加上位概念或下位概念的方法来扩检[4]，若查准率要求较高，应使用专指性较强的概念或增加限制概念来缩小检索范围，还可通过对字段进行限定的方式来保证查找的准确性。对于那些对查全率和查准率无特殊要求的用户来说，也要针对不同的课题，制定相应的检索策略。 对文献量较大或属于成熟学科的课题，应优先考虑查准率，从众多的相关文献中选取针对性较强的文献。对文献较少或新兴学科的课题，可适当放宽检索范围来保证查全率，以免遗漏重要的参考文献。 选择正确的关键词后，就要运用Google检索语法规则构建检索式。 Google无需用明文的“+”来表示逻辑“与”操作，只要空格就可以了。 示例：搜索所有包含关键词“易筋经”和“吸星大法”的中文网页 搜索式：“易筋经吸星大法”（注意：文章中搜索语法外面的引号仅起引用作用，不能带入搜索栏内。） Google用减号“-”表示逻辑“非”操作。

常用搜索引擎的使用方法和技巧

常用搜索引擎的使用方法和技巧 002年12月，教育部出台了《2003～2007年教育振兴行动计划》。从此，教育信息化的观念深入人心。所谓教育信息化，是指在教育领域全面、深入地运用现代信息技术来促进教育改革与发展的过程。其技术特点是数字化、网络化、智能化和多媒体化，基本特征是开放、共享、交互、协作。经过几年的努力，各个学校的信息化建设取得了初步成果，信息化教学环境初具规模。网络是信息化教学环境不可缺少的组成部分，也是重要的信息来源。了解网络搜索引擎技术的基本原理，掌握其应用方法，是教师和学生在信息化教学环境中必备的信息技术素养之一。Google和百度是我们最常用的搜索引擎，一个是外国品牌，一个是民族品牌。下面主要介绍这两种搜索引擎的使用技巧和方法。 Google的使用方法和技巧举例 对于“高级搜索”、“搜索偏好”等选项的使用，大家已经比较熟悉了。在这里，我们介绍一些人们不太注意，但是非常实用的一些功能。 汉语拼音输入检索：为了方便使用中文的用户在网上搜索，Google允许用户直接在键盘上输入汉语拼音来检索相关事物。例如，输入“jisuanji”，检索结果提示：您是不是要找“计算机”？这正是我们需要查找的关键词，用户可以据此浏览相关结果。如果需要查找更详细的资料“联想计算机”，则只要在原来的检索结果“计算机”前输入“lianxiang”。Google的这项新功能，可以免除用户在中文和拼音输入方面的互相转换。用户在输入拼音时，不要留有空格，否则Google会误认为英文。Google会把拼音与常用的字或者词组一一对应。因此，对过于生僻的字或词组，不能用这个方法查找。 语言工具：经常使用计算机的用户手头上自然会有一两个字典软件，用于查找和翻译中英文的词义。Google也提供了一个功能非常强大的语言工具，而且使用很方便。用户输入以下文字：“大量的流行病学调查及多项科研结果显示，雌二醇具有减少脑卒中发作和削减卒中体积的作用。”得到的翻译结果是：“The large number of scientific and epidemiological survey results showed that estradiol with reduced stroke volume and stroke were reduced to the role.” 计算器使用：Google有计算器的功能，例如，在Google检索框中输入“45×86＋35÷7”，就会得到结果：“(45×86)＋(35÷7)=3875”。 检索工具栏：Google的检索工具栏功能强大，有拖放和右击检索功能、新闻阅读、广告拦截、网站排名显示和搜索字词标明等。工具栏可以附在浏览器下，这样使用起来更加方便。用户可以首先在Google网站下载并安装一个检索工具栏，然后根据需要在工具栏的选项中进行设置。这个检索工具栏，能给用户带来许多意想不到的方便。例如：搜索字词标明，通过鲜艳的色彩标明用户所检索的字词在每个网页上的位置，便于用户查阅，单击“搜索字词