Vous êtes sur la page 1sur 3

CSE 446

Lab 1 Report
Name : Nguyen Tien Hieu
Vo Phuc Luan
Dang Ngoc Son
_ Block Diagram:

I( nv pi d u e t eP rs o s ci K _m e nT r i an i tT ie n s uE va at i l (O F ua pl u o t r
o ) n g a n g g o n N o n fa l )
s
+ First we put in video and process them to get featured vector.
+ Then use K-means to divide all featured vector in k group.

+ In training, we create the HMM system by solving the problem 2 of the HMM :
given the observation O = O1O2OT and the model
= (A, B,), how do we choose the corresponding state sequence Q = q1q2qT
which is best explain the observation?

+ In testing, we use the HMM model that we created in the training process to
test the video.
+ Evaluation is to evaluate the system, calculate its accurate.
+ Output: show the result.
_ K_clustering :

+ How K_clustering work:


-

It will randomly choose k center for each group.


For each frame, find out which is the center that has the shortest
distance the frame and assign the group for that frame.
For each group, determine the new center by finding the point that has
the smallest total distance to each point in that group.
Assign the new center to the odd one.
Repeat until the new center and the odd one is the same.

+ K_clustering role in Fall Detection System: is used to separate the feature


vector into k group, each frame will be marked with a number of its group.
_ HMM:
+ How HMM work : is specified by
. N: number of stage
. M: number of possible observations
. The starting state probabilities
P(q0= Si) = ( The probabilities of a stage that it will become the
start stage)
.The state transition probabilities
P(qt+1=Sj | q)
t=Si)=aij
current stage i)

( The probabilities that stage j will be the next stage of

. The observation probabilities


P(Ot=k | qt=Si)=bi(k) ( The probabilities that the observation will be
choose in the i stage)
When we compute all this value we will get the HMM system
+ HMM role Fall Detection System: when we have a set of frames that were
marked with k group. When training for the system, we will put input is a set of
feature vectors and the results of them (0 for nonfall and 1 for fall). Then HMM
will try to find a system that if we put the input like in testing set then we will
have output like the result.
After get the HMM system from the training, for each vector the system will
determine the result of about 20 vectors next to it and put in a buffer. If buffer
have five same result continuously then the stop and get the result for this video.
_ Result in problem 5:

The x-axis is the k number.


The y-axis is the mean distance of all distance from each vector to its center.
I choose the value 100 because it has the smallest mean distance which means
the total distance between the vector and its center is smallest. So the featured
vectors are best divided and the HMM system will be more accurate.

Vous aimerez peut-être aussi