Given a continuous motion planning problem of finding a collision-free path from A to B, the A* search is known to be optimal on a finite sized grid-approximation, where each grid-cell for example has 4 or 8 neighbours.
Now, the A* path might be optimal, but it certainly doesn't have to be exactly the same as the shortest path in continuous space, since the A* path is confined to the grid-cells and integer grid coordinates. Now, I would expect that if you increase the resolution to infinity, the resulting A* path should be exact and equal to the continuous problem. But looking at the literature, I can only find "resolution-completeness" for the cell-based approximation. For example in "A Survey of Motion Planning Algorithms
from the Perspective of Autonomous UAV Guidance" by C. Goerzen, Z. Kong, B. Mettler the Rectanguloid Approximate Cell Decompositon is non-optimal, but resolution-complete. I really don't understand how it's not resolution-optimal, meaning optimal if you increase the resolution to infinity.
My question is whether I am actually interpreting it wrongly or if such a grid-approximation is really never exactly optimal even for infinite resolutions.
Related
I have obtained all the possible paths of maze through image processing. Now, I want to use A* algorithm in order to find shortest path of maze. However, I am confused as to whether euclidean distance will be a better heuristic or manhattan distance. Does it depend upon maze type or is the choice of heuristic independent of maze type? Which distance (manhattan or euclidean) will be a good choice for the following possible paths and why? Please suggest.
PS. (Please add your reference too, if your have any. It will be helpful)
The objective of a Heuristic is to provide contextual information to the pathfinder. The more accurate this information is, the more efficient the pathfinder can be.
You have two contradicting requirements to get a good heuristic, which is good because it means there is a sweet spot. Here they are:
An Heuristic must be admissible, which means it shall never overestimate the distance. Otherwise, the algorithm will be broken and may return paths that are not even optimal.
An heuristic must return the largest distance possible. An heuristic that underestimates the remaining path from a cell, will favour that cell when another might have been better.
Of course the optimal Heuristic would return the exact, correct length (which generally is not achievable or defeats the purpose) because it cannot return a longer path without ceasing to be admissible.
In your case, it looks like you're dealing with 4-connected grids. In that case the manhattan distance will be a better metric than euclidian distance, because the Euclidian will under-estimate the cost of all displacements compared to Manhattan (due to the Pythagorean Theorem).
Whithout any further knowledge than 'the graph is a 4-connected Grid', there is no better metric than Manhattan. If however you manage to obtain more data (obstacle density, 'highways', etc.) then you might be able to devise a better heuristic - though keeping it admissible would be a very hard problem in itself.
EDIT Having a closer look, it looks like you have angled vertices in the bottom left. If that is so, you're not in a 4-connected graph, then you MUST use Euclidian distance, because Manhattan would not be admissible.
Its not clear what moves are available to your hero. Does you graph make up a rectangular grid like chess board and can you go diagonally in one step like king in chess? If yes then Chebyshev distance is the best https://en.wikipedia.org/wiki/Chebyshev_distance.
Otherwise use Euclidian distance.
You cant use Manhattan here if you want an optimal path because Manhattan heuristic is not admissible on diagonal routes (it overestimates them) so it can lead to suboptimal pathes
I am currently working on a experiment that i took multiple photos
of a scene on diferent days with a fixed camera position.
The problem is that on real world it is hard to keep the camera
perfectly fixed.
What i need is to fix the small variance I got automaticaly. The research
I made returned methods considering more complex assumption, like camera
pose estimation, homography estimation etc.
For me its enought to discover just the movement at the image plane returning an
x and y.
A perfect solution would be a function such as:
function [movx movy] = detectMotion(im1,im2).
The solution I already made was to calculate some image features, like harris or
hessian, match them and after manualy select the best ones and use the difference
of their position as a camera displacement estimation. I dont know if this is good
enough but it would be better if it was made automaticaly.
You can do the feature matching automatically be extracting feature descriptors around the interest points. Take a look at this OpenCV tutorial on how to perform feature matching using SURF and FLANN. Once you have the feature matches, run RANSAC or least squares to find the best fit for an x- and y-offset. This will give you a decent estimate of the camera motion.
Another option is to compute sparse optical flow on the detected interest points between the two frames, followed by the RANSAC or least squares procedure as above to compute the best x- and y-offset. Dense optical flow could possibly be more accurate, but at the same time could prove to be overkill.
Let me explain my problem:
I have a black vector shape (let's say it's a series of joined, straight lines for now, but it'd be nice if I could also support quadratic curves).
I also have a rectangle of a predefined width and height. I'm going to place it on top of the black shape, and then take the union of the two.
My first issue is that I don't know how to quickly extract vector unions, but I think there is a well-defined formula I can figure out for myself.
My second, and more tricky issue is how to efficiently detect the position the rectangle needs to be in (i.e., what translation and rotation are needed by the matrices), in order to maximize the black, remaining after the union (see figure, below).
The red outlined shape below is ~33% black; the green is something like 85%; and there are positions for this shape & rectangle wherein either could have 100% coverage.
Obviously, I can brute-force this by trying every translation and rotation value for every point where at least part of the rectangle is touching the black shape, then keep track of the one with the most black coverage. The problem is, I can only try a finite number of positions (and may therefore miss the maximum). Apart from that, it feels very inefficient!
Can you think of a more efficient way of tackling this problem?
Something from my Uni days tells me that a Fourier transform might improve the efficiency here, but I can't figure out how I'd do that with a vector shape!
Three ideas that have promise of being faster and/or more precise than brute force search:
Suppose you have a 3d physics engine. Define a "cone-shaped" surface where the apex is at say (0,0,-1), the black polygon boundary on the z=0 plane with its centroid at the origin, and the cone surface is formed by connecting the apex with semi-infinite rays through the polygon boundary. Think of a party hat turned upside down and crumpled to the shape of the black polygon. Now constrain the rectangle to be parallel to the z=0 plane and initially so high above the cone (large z value) that it's easy to find a place where it's definitely "inside". Then let the rectangle fall downward under gravity, twisting about z and translating in x-y only as it touches the cone, staying inside all the way down until it settles and can't move any farther. The collision detection and force resolution of the physics engine takes care of the complexities. When it settles, it will be in a position of maximal coverage of the black polygon in a local sense. (If it settles with z<0, then coverage is 100%.) For the convex case it's probably a global maximum. To probabilistically improve the result for non-convex cases (like your example), you'd randomize the starting position, dropping the polygon many times, taking the best result. Note you don't really need a full blown physics engine (though they certainly exist in open source). It's enough to use collision resolution to tell you how to rotate and translate the rectangle in a pseudo-physical way as it twists and slides uniformly down the z axis as far as possible.
Different physics model. Suppose the black area is an attractive field generator in 2d following the usual inverse square rule like gravity and magnetism. Now let the rectangle drift in a damping medium responding to this field. It ought to settle with a maximal area overlapping the black area. There are problems with "nulls" like at the center of a donut, but I don't think these can ever be stable equillibria. Can they? The simulation could be easily done by modeling both shapes as particle swarms. Or since the rectangle is a simple shape and you are a physicist, you could come up with a closed form for the integral of attractive force between a point and the rectangle. This way only the black shape needs representation as particles. Come to think of it, if you can come up with a closed form for torque and linear attraction due to two triangles, then you can decompose both shapes with a (e.g. Delaunay) triangulation and get a precise answer. Unfortunately this discussion implies it can't be done analytically. So particle clouds may be the final solution. The good news is that modern processors, particularly GPUs, do very large particle computations with amazing speed. Edit: I implemented this quick and dirty. It works great for convex shapes, but concavities create stable points that aren't what you want. Using the example:
This problem is related to robot path planning. Looking at this literature may turn up some ideas In RPP you have obstacles and a robot and want to find a path the robot can travel while avoiding and/or sliding along them. If the robot is asymmetric and can rotate, then 2d planning is done in a 3d (toroidal) configuration space (C-space) where one dimension is rotation (so closes on itself). The idea is to "grow" the obstacles in C-space while shrinking the robot to a point. Growing the obstacles is achieved by computing Minkowski Differences.) If you decompose all polygons to convex shapes, then there is a simple "edge merge" algorithm for computing the MD.) When the C-space representation is complete, any 1d path that does not pierce the "grown" obstacles corresponds to continuous translation/rotation of the robot in world space that avoids the original obstacles. For your problem the white area is the obstacle and the rectangle is the robot. You're looking for any open point at all. This would correspond to 100% coverage. For the less than 100% case, the C-space would have to be a function on 3d that reflects how "bad" the intersection of the robot is with the obstacle rather than just a binary value. You're looking for the least bad point. C-space representation is an open research topic. An octree might work here.
Lots of details to think through in both cases, and they may not pan out at all, but at least these are frameworks to think more about the problem. The physics idea is a bit like using simulated spring systems to do graph layout, which has been very successful.
I don't believe it is possible to find the precise maximum for this problem, so you will need to make do with an approximation.
You could potentially render the vector image into a bitmap and use Haar features for this - they provide a very quick O(1) way of calculating the average colour of a rectangular region.
You'd still need to perform this multiple times for different rotations and positions, but it would bring it algorithmic complexity down from a naive O(n^5) to O(n^3) which may be acceptably fast. (with n here being the size of the different degrees of freedom you are scanning)
Have you thought to keep track of the remaining white space inside the blocks with something like if whitespace !== 0?
What are some path finding algorithms used in games of all types? (Of all types where characters move, anyway) Is Dijkstra's ever used? I'm not really looking to code anything; just doing some research, though if you paste pseudocode or something, that would be fine (I can understand Java and C++).
I know A* is like THE algorithm to use in 2D games. That's great and all, but what about 2D games that are not grid-based? Things like Age of Empires, or Link's Awakening. There aren't distinct square spaces to navigate to, so what do they do?
What do 3D games do? I've read this thingy http://www.ai-blog.net/archives/000152.html, which I hear is a great authority on the subject, but it doesn't really explain HOW, once the meshes are set, the path finding is done. IF A* is what they use, then how is something like that done in a 3D environment? And how exactly do the splines work for rounding corners?
Dijkstra's algorithm calculates the shortest path to all nodes in a graph that are reachable from the starting position. For your average modern game, that would be both unnecessary and incredibly expensive.
You make a distinction between 2D and 3D, but it's worth noting that for any graph-based algorithm, the number of dimensions of your search space doesn't make a difference. The web page you linked to discusses waypoint graphs and navigation meshes; both are graph-based and could in principle work in any number of dimensions. Although there are no "distinct square spaces to move to", there are discrete "slots" in the space that the AI can move to and which have been carefully layed out by the game designers.
Concluding, A* is actually THE algorithm to use in 3D games just as much as in 2D games. Let's see how A* works:
At the start, you know the coordinates of your current position and
your target position. You make an optimistic estimate of the
distance to your destination, for example the length of the straight
line between the start position and the target.
Consider the adjacent nodes in the graph. If one of them is your
target (or contains it, in case of a navigation mesh), you're done.
For each adjacent node (in the case of a navigation mesh, this could
be the geometric center of the polygon or some other kind of
midpoint), estimate the associated cost of traveling along there as the
sum of two measures: the length of the path you'd have traveled so
far, and another optimistic estimate of the distance that would still
have to be covered.
Sort your options from the previous step by their estimated cost
together with all options that you've considered before, and pick
the option with the lowest estimated cost. Repeat from step 2.
There are some details I haven't discussed here, but this should be enough to see how A* is basically independent of the number of dimensions of your space. You should also be able to see why this works for continous spaces.
There are some closely related algorithms that deal with certain problems in the standard A* search. For example recursive best-first search (RBFS) and simplified memory-bounded A* (SMA*) require less memory, while learning real-time A* (LRTA*) allows the agent to move before a full path has been computed. I don't know whether these algorithms are actually used in current games.
As for the rounding of corners, this can be done either with distance lines (where corners are replaced by circular arcs), or with any kind of spline function for full-path smoothing.
In addition, algorithms are possible that rely on a gradient over the search space (where each point in space is associated with a value), rather than a graph. These are probably not applied in most games because they take more time and memory, but might be interesting to know about anyway. Examples include various hill-climbing algorithms (which are real-time by default) and potential field methods.
Methods to procedurally obtain a graph from a continuous space exist as well, for example cell decomposition, Voronoi skeletonization and probabilistic roadmap skeletonization. The former would produce something compatible with a navigation mesh (though it might be hard to make it equally efficient as a hand-crafted navigation mesh) while the latter two produce results that will be more like waypoint graphs. All of these, as well as potential field methods and A* search, are relevant to robotics.
Sources:
Artificial Intelligence: A Modern Approach, 2nd edition
Introduction to The Design and Analysis of Algorithms, 2nd edition
I'm looking for a way to determine the optimal X/Y/Z rotation of a set of vertices for rendering (using the X/Y coordinates, ignoring Z) on a 2D canvas.
I've had a couple of ideas, one being pure brute-force involving performing a 3-dimensional loop ranging from 0..359 (either in steps of 1 or more, depending on results/speed requirements) on the set of vertices, measuring the difference between the min/max on both X/Y axis, storing the highest results/rotation pairs and using the most effective pair.
The second idea would be to determine the two points with the greatest distance between them in Euclidean distance, calculate the angle required to rotate the 'path' between these two points to lay along the X axis (again, we're ignoring the Z axis, so the depth within the result would not matter) and then repeating several times. The problem I can see with this is first by repeating it we may be overriding our previous rotation with a new rotation, and that the original/subsequent rotation may not neccesarily result in the greatest 2D area used. The second issue being if we use a single iteration, then the same problem occurs - the two points furthest apart may not have other poitns aligned along the same 'path', and as such we will probably not get an optimal rotation for a 2D project.
Using the second idea, perhaps using the first say 3 iterations, storing the required rotation angle, and averaging across the 3 would return a more accurate result, as it is taking into account not just a single rotation but the top 3 'pairs'.
Please, rip these ideas apart, give insight of your own. I'm intreaged to see what solutions you all may have, or algorithms unknown to me you may quote.
I would compute the principal axes of inertia, and take the axis vector v with highest corresponding moment. I would then rotate the vertices to align v with the z-axis. Let me know if you want more details about how to go about this.
Intuitively, this finds the axis about which it's hardest to rotate the points, ie, around which the vertices are the most "spread out".
Without a concrete definition of what you consider optimal, it's impossible to say how well this method performs. However, it has a few desirable properties:
If the vertices are coplanar, this method is optimal in that it will always align that plane with the x-y plane.
If the vertices are arranged into a rectangular box, the box's shortest dimension gets aligned to the z-axis.
EDIT: Here's more detailed information about how to implement this approach.
First, assign a mass to each vertex. I'll discuss options for how to do this below.
Next, compute the center of mass of your set of vertices. Then translate all of your vertices by -1 times the center of mass, so that the new center of mass is now (0,0,0).
Compute the moment of inertia tensor. This is a 3x3 matrix whose entries are given by formulas you can find on Wikipedia. The formulas depend only on the vertex positions and the masses you assigned them.
Now you need to diagonalize the inertia tensor. Since it is symmetric positive-definite, it is possible to do this by finding its eigenvectors and eigenvalues. Unfortunately, numerical algorithms for finding these tend to be complicated; the most direct approach requires finding the roots of a cubic polynomial. However finding the eigenvalues and eigenvectors of a matrix is an extremely common problem and any linear algebra package worth its salt will come with code that can do this for you (for example, the open-source linear algebra package Eigen has SelfAdjointEigenSolver.) You might also be able to find lighter-weight code specialized to the 3x3 case on the Internet.
You now have three eigenvectors and their corresponding eigenvalues. These eigenvalues will be positive. Take the eigenvector corresponding to the largest eigenvalue; this vector points in the direction of your new z-axis.
Now, about the choice of mass. The simplest thing to do is to give all vertices a mass of 1. If all you have is a cloud of points, this is probably a good solution.
You could also set each star's mass to be its real-world mass, if you have access to that data. If you do this, the z-axis you compute will also be the axis about which the star system is (most likely) rotating.
This answer is intended to be valid only for convex polyhedra.
In http://203.208.166.84/masudhasan/cgta_silhouette.pdf you can find
"In this paper, we study how to select view points of convex polyhedra such that the silhouette satisfies certain properties. Specifically, we give algorithms to find all projections of a convex polyhedron such that a given set of edges, faces and/or vertices appear on the silhouette."
The paper is an in-depth analysis of the properties and algorithms of polyhedra projections. But it is not easy to follow, I should admit.
With that algorithm at hand, your problem is combinatorics: select all sets of possible vertexes, check whether or not exist a projection for each set, and if it does exists, calculate the area of the convex hull of the silhouette.
You did not provide the approx number of vertex. But as always, a combinatorial solution is not recommended for unbounded (aka big) quantities.