Artificial Intelligence DD2380

HMM example: Usain Bolt

You have developed an automatic video annotation system for annotating recorded running sequences of Usain
Bolt. The system takes images from a video stream of a running sequence as an input, it extracts some visual data
and annotates each image as:

• Usain is preparing for running

• Usain runs/accelerates
• Usain decelerates
You have built the following model for the system:

0.3 0.9 0.4

0.7 0.1 0.6

PREPARE ACCELERATE DECELERATE STOP

F1=0.5 F3=0.2 F4=0.1

F2=0.2 F4=0.7 F6=0.5
F3=0.3 F5=0.1 F7=0.4

Here, F 1 . . . F 7 represent seven ways of classifying each video frame (i.e. seven di↵erent possible observation
outcomes that might have to do with for example Usain’s posture) and their probabilities of being observed in each
state. If an observation is not mentioned under a state it means it has probability zero of being observed there. The
states are annotated by xt and observations in each time step are represented by ot . Note that STOP is a non-emitting
terminating (or absorbing) state.
The system divides a video sequence in several video frames and each of them is classified to one of the F 1 . . . F 7.

1. What are A, B and ⇡?

2. What is P (x2 = ACCELERAT E)?
3. What is P (o2 = F 2)?
4. What is P (x2 = ACCELERAT E|o2 = F 2)?

5. If we assume o1 = F 1, o2 = F 3, o3 = F 4, o4 = F 5, o5 = F 6. What is the most likely o6 ?

6. If we assume o1 = F 1, o2 = F 2, o3 = F 3, o4 = F 4, o5 = F 4, o6 = F 6, o7 = F 7. What is the most likely
sequence of states?

1
 Artificial Intelligence DD2380
HMM example: Weather prediction

Your housekeeping robot hands you the right coat each day before you leave for work. The robot cannot see the
weather conditions, but it does have a crystal which reacts to environmental conditions. The crystal turns into one
of 3 colors (red, green or blue). That is, the actual weather conditions (sun, rain, snow) are hidden to the robot. The
Hidden Markov Model for the problem is given below with state transition and observation matrices.

State transitions Observations

SUN RAIN SNOW RED GREEN BLUE
SUN 0.7 0.2 0.1 SUN 0.75 0.15 0.1
RAIN 0.3 0.5 0.2 RAIN 0.35 0.5 0.15
SNOW 0.2 0.3 0.5 SNOW 0.25 0.2 0.55

1. Start by drawing a model of the system with hidden and observable states annotated. Do not forget to draw all
the connections between the hidden states, as well as between states and observations.
2. You do not know the exact initial state of the world but you know from previous experience that it is a about a
50% chance of sun and equal chance for rain and snow. We have O = {o1 = blue, o2 = green, o3 = red}.

3. What is the most likely weather condition for the forth day?
4. What is the most likely observation for the fourth day?
5. What is the most likely weather condition each of the three days first days?