I try to use Dynamic Time Warping (DTW) to detect gestures performed with a smartphone by using the accelerometer sensor. I already implemented a simple DTW-algorithm.
So basicly I am comparing arrays of accelerometer-data (x,y,z) with DTW. The one array contains my predefiend gesture, the other should contain the measured values. My problem is, that the accelerometer-sensor measures continously new values and I don't know when to start the comparison with my predefined value-sequence.
I would need to know when the gesture starts and when it ends, but this might be different with different gestures. In my case all supported gestures start and end at the same point, but as far as I know I can't calculate the traveled distance from acceleration reliably.
So to sum things up: How would you determine the right time to compare my arrays using DTW?
Thanks in advance!
The answer is, you compare your predefined gesture to EVERY
subsequence.
You can do this in much faster than real time (see [a]).
You need to z-normalize EVERY subsequence, and z-normalize your predefined gesture.
So, by analogy, if you stream was.....
NOW IS THE WINTER OF OUR DISCONTENT, MADE GLORIOUS SUMMER..
And your predefined word was made, you can compare with every marked word beginning (denoted by white space)
DTW(MADE,NOW)
DTW(MADE,IS)
DTW(MADE,THE)
DTW(MADE,WINTER)
etc
In your case, you don’t have makers, you have this...
NOWISTHEWINTEROFOURDISCONTENTMADEGLORIOUSSUMMER..
So you just test every offset
DTW(MADE,NOWI)
DTW(MADE, OWIS)
DTW(MADE, WIST)
DTW(MADE, ISTH)
::
DTW(MADE, TMAD)
DTW(MADE, MADE) // Success!
eamonn
[a] https://www.youtube.com/watch?v=d_qLzMMuVQg
You want to apply DTW not only to a time-series, but to a continously evolving stream. Therefore you will have to use a sliding window of n recent data points.
This is exactly, what eamonn described in his second example. His target pattern consists of 4 events (M,A,D,E) and therefore he uses a sliding window with length of 4.
Yet in this case, he makes the assumption, that the data stream contains no distortions, such as (M,A,A,D,E). The advantage of DTW is that it allows these kind of distortions and yet recognizes the distorted target pattern as a match. In your case, distortions in time are likely to happen. I assume that you want equal gestures performed either slow or fast as the same gesture.
Thus, the length of the sliding window must be higher than the length of the target pattern (to be able to detect a slow target gesture). This is computationally expensive.
Finally, my point is: I want to recommed you this paper
Spring algorithm by Sakurai, Faloutsos and Yamamuro.
They optimized the DTW algorithm for datastreams. You will no longer need more than n*n computations per incoming event but only n. It basically is DTW but cutting down all unneccesary computations and only taking the best possible alignment of the template onto the stream into account.
p.s. most of what I know about time-series and pattern matching, I learned by reading what Eamonn Keogh provided. Thanks a lot, Mr. Keogh.
Related
My device has only two focus modes, AUTO and FIXED (as per getSupportedFocusModes()).
I want to set my camera at a fixed focus distance of 'x' (x being whatever I like, or whatever I can get from the camera..). (I'm aware of setFocusMode(Camera.Parameters.FOCUS_MODE_FIXED), but this seems to be fixed only on the farthest possible setting..)
Can this be done? (Android version 4.2.2)
Not trying to completely answer the question here, just trying to give it some direction.
So, what you need here is a driver support for that kind of operation. Then at some point you'd ask the driver from your application to set a requested focus distance.
Another question is: "if anyone really needs that kind of functionality?".
Android documentation says:
public static final String FOCUS_MODE_FIXED
Focus is fixed. The camera is always in this mode if the focus is not adjustable. If the camera has auto-focus, this mode can fix the focus, which is usually at hyperfocal distance. Applications should not call autoFocus(AutoFocusCallback) in this mode.
Lets see what hyperfocal distance is.
Hyperfocal distance
From Wikipedia, the free encyclopedia
In optics and photography, hyperfocal distance is a distance beyond which all objects can be brought into an "acceptable" focus. There are two commonly used definitions of hyperfocal distance, leading to values that differ only slightly:
Definition 1: The hyperfocal distance is the closest distance at which a lens can be focused while keeping objects at infinity acceptably sharp. When the lens is focused at this distance, all objects at distances from half of the hyperfocal distance out to infinity will be acceptably sharp.
Definition 2: The hyperfocal distance is the distance beyond which all objects are acceptably sharp, for a lens focused at infinity.
The distinction between the two meanings is rarely made, since they have almost identical values. The value computed according to the first definition exceeds that from the second by just one focal length.
As the hyperfocal distance is the focus distance giving the maximum depth of field, it is the most desirable distance to set the focus of a fixed-focus camera.
So the focus is not set on the farthest possible setting, but is set to have all visible objects to be acceptably sharp.
Returning to the question.
If you happen to be a developer of this particular camera's firmware, you can add any needed IOCTLs to you driver. But then you still going to need to call them somehow. This can't be achieved without adding additional functions into the Android OS, and further recompiling of Android itself and it's underlying Linux kernel.
So it seems like you can't achieve this goal, not from the user space at least.
One potential approach to achieve that fixed focus distance is to call autoFocus at the start of the camera life-cycle. Keep calling autoFocus sporatically until a condition is met. Once the condition is met, then instead of calling autoFocus, set a flag and call takePicture instead.
This is one solution that I have come to in order to get the desired effect that you might be looking to achieve.
So within my thread that is taking pictures continuously, the code looks something like this:
if(needsFocus)
{
myCamera.autoFocus(autoFocusCallback);
}
else //Focus is not needed anymore at this point
{
if(myCamera != null)
{
myCamera.startPreview();
myCamera.takePicture(pictureCallback);
}
}
Once the condition is met, needsFocus is set to true. At this point, the focus is fixed at the place that I want it to be at. Then it won't change throughout the rest of the activities task. The condition for my case was the appearance of a particular object detected with the OpenCV library.
I might be wrong, but the way you phrase your question seems like coming from a classic DSLR lens perspective.
On an android mobile camera, you don't actually have to worry that much of a lens focal distance, unless your mobile camera allows that (which does not seem to be the case, as you mention it just allows auto or fixed, instead of infinite, macro, continuous-video, etc).
You can just set local areas on the camera to focus and let the sdk do its work. If the object touched on the camera image is far or near it's the sdk work to calculate accordingly and focus for you.
For an example, try this open-source camera project.
I need to generate objects for my little character to jump over. I have the artwork for these obsticles and my character can jump and I have a scrolling background.
How can I spawn my artwork for my obsticles on my x axis with spacing inbetween them?
Can anyone provide me with some sample code or atleast try and point me in the right direction?
Many Thanks,
James
Yes. You can. You want to use some sort of loop that generates them:
2 options you can use:
local function frameHandler()
if should_I_make_object() then
createObstacle()
end
end
Runtime:addEventListener("enterFrame", frameHandler)
this approach will create new objects according to frame rate. IE, lets
you create objects every 100 frames lets say. This will make levels play the
same (have the same number of obstacles) on different devices that have varied frame-rate
Option 2:
local function createObstacle()
--your_create_obstacle_code()
if game_is_still_playing() then
timer.performWithDelay(object_spawn_delay, createObstacle)
end
end
This option will create a new object every object_spawn_delay milliseconds.
This is easy to code, and is a nice solution when you need things to happen on
a time-dependent interval. But you do need code to decide if the game is still playing.
Also, be aware that if the game ends, there might still be a lingering callback to
createObstacle() that can create nasty bugs. Make sure to do proper cleanup when
the level / game ends and be aware this callback may be an issue.
I see queries related to opencv motion detection, but my requirement is much simpler , so i am asking the question again .
I would like to analyse video frames and see if something has changed in the frame. Any kind of motion occurring in the frame has be recognized. I just want to get notified if something happens. I don't need to track/ draw contours.
Attempts made :
1) Template matching using OpenCV ( TM_CCORR_NORMED ).
I get the similarity index using cvMinMaxLoc &
if( sim_index > threshold )
"Nothing chnged"
else
"Changed
Problem faced :
I couldn't find a way to decide on how to set thresholds. The values of false match and perfect were very close.
2) Method 2
a) Make running average
b) Take abs difference between current frame and moving average.
c) Threshold it and made it binary
d) Count the number of non zero values
Again am stuck with how to threshold it, because i am getting a large number of non zero values even for very similar frames.
Please advice me on what approach i should take. Am i going in the right direction with the above two methods, or is there a simple method which can work in all most generic scenarios.
Method 2 is generally regarded as the most simple method for motion detection, and is very effective if you have no water, swaying trees or highly variable lighting conditions in your video.
Normally you implement it like this:
motion_frame=abs(newframe-running_avg);
running_avg=(1-alpha)*running_avg+alpha*newframe;
You can threshold the motion_frame if you want, then count the nonzeroes. But you could also just sum the elements of the motion_frame and threshold that instead (be sure to work with floating point numbers). Optimizing the parameters for this is pretty easy, just make two trackbars and play around with it. Typically alpha is around [0.1; 0.3].
Lastly, it is probably overkill to do this on entire frames, you could just use subsampled versions and the result will be very similar.
I'm writing this game on Android where I have a bunch of characters moving around who collide with each other. Everything works fine but when I get passed a certain number of characters on the screen at the same time, the performance of the app gets hit severely. I did my tests and drawing is not causing the low frame rate, it is the algorithm for collision detection, since every time they move they have to check their location to all the other characters. So currently I'm just looping through them all for each character. Is there a way to improve on this? Is there a performance trick to collision detection on a big number of objects that I don't know about?
Yes, there is a technique based on a first broad-phase and second narrow-phase colission detection.
I'll quote some paragraps from: Beginning Android Games, by Mario Zechner.
Broad phase: In this phase we try to figure out which objects can
potentially collide. Imagine having 100 objects that could each
collide with each other. We’d need to perform 100 * 100 / 2 overlap
tests if we chose to naively test each object against each other
object. This naive overlap testing approach is of O(n^2) asymptotic
complexity, meaning it would take n^2 steps to complete (it actually
finished in half that many steps, but the asymptotic complexity
leaves out any constants). In a good, non-brute-force broad phase, we
try to figure out which pairs of objects are actually in danger of
colliding. Other pairs (e .g., two objects that are too far apart for
a collision to happen) will not be checked . We can reduce the
computational load this way, as narrow-phase testing is usually pretty
expensive.
Narrow phase: Once we know which pairs of objects can potentially
collide, we test whether they really collide or not by doi ng an
overlap test of their bounding shapes.
The broad phase involves dividing the world in large cells, making some sort of grid.
Each cell has the exact same size, and the whole world is covered in cells. If two objects are not in the same cell, a narrow phase for those two objects is not needed.
Quote once again:
All we need to do is the following:
Update all objects in the world based on our physics and controller step.
Update the position of each bounding shape of each object according to the object’s position. We can of course also include the orientation and scale as well here.
Figure out which cell or cells each object is contained in based on its bounding shape, and add it to the list of objects contained in those cells.
Check for collisions, but only between object pairs that can collide (e.g., Goombas don’t collide with other Goombas) and are in the same cell.
This is called a spatial hash grid broad phase, and it is very easy to implement. The first thing we have to define is the size of each cell. This is highly dependent on the scale and units we use for our game’s world.
It also depends on the bounding shape you're using. A simple rectangle or circle around the characters and it's euclidean distance is one simple thing to calculate, but a finer shape (including details as "the head", "the legs" with little additional bounding shapes) will be more a lot more computationally expensive to calculate.
If all objects are free to move to any part of the screen, then the best you can do is your O(n^2) algorithm. You can improve it by a constant factor by realizing that when you check if object A collides with object B, then you don't have to later check if object B collides with object A.
enclose each character within a fixed size square. Before you check for character collision, check if the squares in which they are enclosed collide. If and only if the squares collide, there would be a chance for the characters to collide. Now checking for squares collision is easy as you have to just compare the x & y co-ordinates.
Dividing into a broad phase and narrow phase as Federico suggests only helps if your collision detection algorithm is expensive, i.e. it's not a simple bounding box.
Fortunately there are other options.
You could try a collision mask technique. Since you don't seem to be limited by rendering speed, render a bounding box for each object into a hidden bitmap. Before rendering the next object, check the pixels at the four corners of its bounding box to see if they have already been written. You can even use a different colour for each object so that the colour tells you which object the collision was with.
Another popular trick is to simply not do every collision check every frame. For example, games like Super Mario Bros actually only check for collisions between the player and enemies every other frame. You can do a more advanced version where you check all objects in a round-robin fashion, doing as many as it can per frame. When things get busy each object might only be checked every other or even every third frame, but the player is unlikely to notice. This works best if your objects are not moving so fast that they can pass through each other one only one frame of collision.
So in my game my View gets drawn an inconsistent rates. Which in turn makes it glitchy. Ive been running into alot of problems with the invalidate(); meathod. Any simple ideas- Everywhere i look I get thrown up on by tons of intense code.
You haven't provided us with much information, specifically code.
A few things you could do are:
Set the initial frame rate to the lowest value you observe your application runs at, i.e., if currently set to 1/60, but the frame rate continuously dips to 1/30, set to 1/30 etc.
Rework your drawing calls to be more efficient.
Try to combine multiple transformations into a single transformation by multiplying matrices, i.e. if you need to scale, translate, and rotate, multiply those three matrices together and apply that single transformation to the vertices instead of applying three separate transformations.
Try not to iterate through entire lists/arrays if unnecessary.
Attempt to use the lowest level / most primitive structure possible for anything you have to process in the loop to avoid the overhead of unboxing.
[edit on 2012-08-27]
helpful link for fixing you timestep: http://gafferongames.com/game-physics/fix-your-timestep/
It sounds like your game loop doesn't take into account the actual time that has passed between iterations.
The problem is the assumption that there is a fixed amount of time between loop iterations. But this time can be variable depending on the number of objects in the scene, other processes on the computer, or even the computer itself.
This is a common, somewhat subtle, mistake in game programming, but it can easily be remedied. The trick is to store the time at the end of each draw loop and then take the difference of the last update with the current time at the start. Then you should scale all animations and game changes based on the actual elapsed time.
I've wrote more about this on my blog a while back here: http://laststop.spaceislimited.org/2008/05/17/programming-a-pong-clone-in-c-and-opengl-part-i/
Part II specifically covers this issue:
http://laststop.spaceislimited.org/2008/06/02/programming-pong-in-c-and-opengl-part-ii/