Published Indoor Navigation
Published Indoor Navigation
ABSTRACT
We have tried to analyze the indoor navigation of a small vehicle for which we have used a Data set with time
series. The data is collected as Radio Signals. This is an intuitive approach and by understanding it deeply we
will make navigation better.
INTRODUCTION:
We have imported the dataset from the UCI repository. We used the dataset of ambient assisted Living.
Though, originally the data is collected by predicting movements of a user within an office, we have tried to fit
it as per our convenience. As an input, we have Radio Signals. It consists of a time series that is collected from
a network of Wireless sensors. Taking one room into consideration, we will have four signal emanators fixed
at four corners of the room which will serve as anchors. Then the fifth sensor will be fixed on the small vehicle,
say a car. The data is collected at a frequency of 5Hz. We then convert the data to a scale between [0,1], i.e.
rescaling. We solely use the data coming from each of the anchors only. Then, we have the target data which is
acts like an inference table. It has a labeled class that will indicate whether the vehicle’s trajectory will change
spatially or not. The entries alternate between -1 and +1. The class +1 relates to change of trajectory and
Movement and the class -1relatestocontinuinginthesamepathi.e.preserving.
networks is its ability to self-learn. Varied functions like summation where all input values are added up before
transmitting it to a new nerve unit.
MODEL CHOICE:
Totally dependent on the representation of data along with its varied applications. The more complex the
model, the more challenging it is Algorithm used: There are many algorithms for learning and will work
accurately with the training of a fixed dataset and correct parameters.
_ Durability: With appropriate selection of model, cost function and algorithm used for learning, it can
make the entire neural network powerful.
B1 Recurrent Neural Network:
There are various classes under Artificial Neural Networks. The one that fits our problem statement is the
Recurrent Neural Network. Here the various connections between the neural units form a directed cycle. An
internal network state is created and it thereby exhibits temporal behavior. This internal memory of the
networks are used to process the random series of inputs math.
APPLICATIONS:
Enlisted below are some of the applications of these networks:
Approximation of functions, regression, time series and modelling. E.g.: radio signals prediction for
temporal data like ours.
Recognition of patterns and sequences that help in classification. It also helps in collision detection
etc.
Processing of data using various technques like filtering, clustering, compression etc.
OUR METHODOLOGY: We have used the dataset that has records of radio signal strengths. The
propagation or movement of our vehicle will have the following phases. Forward movement of Vehicle in the
training set’s input through the network system for activation of its output. Using training values the vehicle can
also have backward movement through the network systems. Then it helps generate the difference between the
actual and predicted values(delta) and accordingly gives a visual representation in the form of a plot.Where we
multiply the delta value of output and activation value of the input to obtain the weights’ gradient. We then
subtract the ratio from the gradient. The quality and accuracy of learning is influenced by the ratio obtained. The
bigger the ratio, the faster the nerve unit trains, but with a lower ratio the Accuracy of the training is increased.
So by constantly repeating the above-mentioned process we can achieve the desired performance.
CODE:
>install.packages("rnn")
>library(rnn)
>require(rnn)
>set.seed(10)
> f <- 5
> w <- 2*pi*f
> x <- sin(t*w) + rnorm(200, 0, 0.25)
> y <- cos(t*w)
> X <- matrix(X, nrow =300)
> Y <- matrix(Y, nrow =300)
The package is installed. The obtained radio signals can be visually represented as a sine wave
with necessary modifications in its electric and magnetic field.So here, we have set the frequency
as 5.The seed value helps us in prediction.These signals are normalized and the waves are fit in the matrix
form for our dataset.
plots
A regular signal has number of disturbances and that gives us the noisy waves. >plot(as.vector(X),
col=’blue’, type=’l’, ylab = "Signals", main = "Noisy waves") lines(as.vector(Y), col = "red")
legend("topright", c("X", "Y"), col = c("blue","red"), lty = c(1,1), lwd = c(1,1))
The dataset is divided into train and test which enables us in comparing the reults and reaching
an effective conclusion.
>train <- 1:2
>test <- 3:4
>model <- trainr(Y = Y[train,], X = X[train,],learninggrate = 0.05,hiddendim = 16,numepoch =
1500)
>Yp <- predictr(model, X)
>plot(as.vector(t(Y[test,])), col = ’red’, type=’l’, main = "Actual vs predicted: testing set", ylab
= "Y,Yp")
Target dataset:
So, suppose the vehicle wants to travel from A to B. There can be several paths to reach the destination
B.B ut our vehicle will solely move in the direction where the signal strength is highest. The various paths are
pre-fed in the vehicle s memory. So what happens if the path to reach B is north, and the signal closer to the
vehicle is in the direction of south? Well, then according to our analysis, it will check the pre-fed path. So, if
there is a change in the direction, it is denotedby+1andpreservingofthedirectionisdenotedby-1.0
CONCLUSION:
There are a number of models that can be fitted for this problem statement like the LSTM-Long Short term
Memory which is also a type of Recurrent Neural Network. We have used the Plain Vanilla Recurrent Model
as a simple intuitive approach for the indoor navigation of a small vehicle and thus this model can further be
improvised using in-set technologies.
Article
Ralf Solomon, J.L.van.Hemman-'Article on accelerating backpropogation through self-adaptation'