My project
I was thinking about an AI that could improve the everyday life of seniors. Then, I though to how improve their security in at home. For instance, seniors who don't have the chance to have friends or relatives close to their home could be in danger and nobody could notice the problem. In this perspective, I though to an AI which could be able to react if seniors fall down. The AI detects if the senior fell on the ground, I mean an unusual movment, if it is the case a alarm could automatically sound and emergency would be aware of this fall.
The AI would make difference between a serious fall and just a little fall. If it's a serious fall the alarm would be more intense and a specific message could appear, on the contrary if the AI detects a fall but then the senior could stand up, in this case the alarm would be less intense and a message which informs emergency that they should visit the senior could appear.
STEP 1 - The Gesture recognition
The First Step, is the implementation of the AI. In this stage, we have to make the Ai learn when a person fall on the ground. For this, I propose to make an AI learning with thousand of videos about person who fall.
That is to say, we have to integrate thousands videos that the AI could analyse. For instance the AI could focus on particular and unusual gesture, which are translated by ususual position of the key point recognition. This recognition could be made with the position of the 17 key points of the body.
When a person fall on the ground, those 17 key points have a specific location. For instance, the arm is at the same level as the knee, and those two points are really close to the ground. Otherwise, if the old people stay on the floor without any movment, the AI could detect that gesture as an usnusual one.
To do that, it exists a ml5 tool which is able to recognize those movments. The following code use Pose Net and ml5.js. You can find it entirely Here.
In the code you just have to replace the "Default Values" by the ones of a body position of a person who fell down. Obvsiouly it will be better to take many different values because of the multiple position how you could fall.
You can find the entire code here!
After the machine learned how a person fall and the ususual gesture or position that have been adopted. We could integrate a difference between a serious fall or a simple one, for instance the could recognize if during the minute following the fall the person has bee, able to stand up. If it's the case the AI should alarm the emergency with a message less alarming and in the other case, it should send an emergency message while precising it has to be taken seriously.
Let's now move on the second step where the AI is gonna act!!
When a person fall on the ground, those 17 key points have a specific location. For instance, the arm is at the same level as the knee, and those two points are really close to the ground. Otherwise, if the old people stay on the floor without any movment, the AI could detect that gesture as an usnusual one.
To do that, it exists a ml5 tool which is able to recognize those movments. The following code use Pose Net and ml5.js. You can find it entirely Here.
In the code you just have to replace the "Default Values" by the ones of a body position of a person who fell down. Obvsiouly it will be better to take many different values because of the multiple position how you could fall.
You can find the entire code here!
After the machine learned how a person fall and the ususual gesture or position that have been adopted. We could integrate a difference between a serious fall or a simple one, for instance the could recognize if during the minute following the fall the person has bee, able to stand up. If it's the case the AI should alarm the emergency with a message less alarming and in the other case, it should send an emergency message while precising it has to be taken seriously.
Let's now move on the second step where the AI is gonna act!!
STEP 2 - Understand if it's a serious fall
IF IT'S YES, THEN...
Now it would be enought to send a message to the nearest doctor. When the AI recognize the caracteristic position of a fallen person, it determines if the person is able to move or not (We can put a timer of one minute, if this one minute the AI records no moves of the 17 key points, then AI is gonna act). If it's not, the AI is gonna send an alarming message to all the doctor of the neighborhood until one answers that it takes situation under control (I mean almost like the UBER reallocation).
IF IT'S NOT, THEN THE AI CHANGE THE MESSAGE...
So, we are in the case of a not serious fall, I mean a fall but the person could stand up after and make himself the necessary measure (call relatives or friends). In this situation, the AI is not gonna send the same alarming message. The doctor would be aware of the fall but it would be precise that the person could stand up. So the type of intervention could change. Doctors just visit the person to check if everything is correct or if a medical assistance is necessary.
STEP 3 - Transfer the information to Emergency
Now we are in the final stage, it remains to define two different types of message and send them to doctors.
To do that, it could be necessary to code specific reaction related to the recorded movments. This reaction would be divided into the two available message that we have seen above. You can find here a special code to test the program!
To do that, it could be necessary to code specific reaction related to the recorded movments. This reaction would be divided into the two available message that we have seen above. You can find here a special code to test the program!