Status Report #11
Joel:
The last part of our project was to make sure that there were no false positives from our neural network. This would be a bit difficult since we only trained our network to recognize the 7 of us. So in order for it to determine if someone wasn’t part of the seven residents, we did 3 things. The first thing we did was to add repetition when calculating who is at the door. So this meant checking twice if the person at the door indeed equated to the same output from the neural network and wasn’t just a random fluke. The second thing we did was to actually implement 2 neural networks. One was a bit higher and accuracy and took longer to run, while the other was faster and had a different activation function on the output layer. So we inputted the person’s face who was at the door two both of these networks and wanted to make sure that both networks identified them as the same person. The last thing we did to make sure we didn’t have any false positives was to have a threshold of accuracy for our neural network outputs. So each neural network would either output a percentage of how sure it was that the person at the door was each of the 7 residents, and the other would output weights instead of percentages. I averaged the weights and percentages for our test data set, found the mean of every number, reduced a bit, and then set that as the threshold needed from each network in order to determine if that was a resident.
Chinedu:
In this final week of Capstone, we switched gears towards final circuit assembly and building the door itself and the means to house all of our hardware. This entailed first placing attaching the circuit to the system using tape in order to verify how the system will be composed structurally and to test the frames and the distances that change how our algorithm would be run. After this, we went to the construction phase in which we took our door to the maker space to build a stand for the door, attach the strike to the door, and build the holes to house our circuit, keypad, and camera. This was a different experience from my usual ECE coursework, but was a pleasant change from the usual coding of computer engineering. To do this I had to completely remake the circuit to work using the dimensions of our door. I had to make the circuit more concise and to have a smaller physical footprint.
Omar:
This was the final week, so things started getting real. We had a lot of physical work to do on the door, and not much time to work with. With three days left, we had made appropriate holes in the door, We started by cutting out a rectangular panel in the back of the door and inserting all our electrical components, from the breadboard which contained the main logic of the circuit, to the Raspberry Pi, which all fit conveniently in a housing I created. I made the housing in the Makerspace; after dimensioning the holes we had already made, it didn’t take long to design a housing for some electrical parts we had. I also designed a housing for the ultrasonic sensor/Pi camera duo at the front hole of the door. This made for a flush outside look for the door, which came in clutch for the last few days.
For the door, we had to do a lot of mechanical engineering work to keep the door upright. After some time deliberating a design to keep the door standing by itself, we set to work by getting a wooden pallet and cutting out the wooden parts we needed. We then assembled a structure with 4 legs and triangle supports that made the door pretty difficult to move, but kept it upright.