Carnegie Mellon ECE Capstone // Spring 2024 // Ankita Chatterjee, Kaitlyn Liu, Zina Zarzycki
This ended up being a very busy week for me, so I got a bit behind on my scheduled deliverables. I did create a lot clips of traffic camera footage using live camera feeds online so that Ankita could begin testing the object detection code. …
Potential Risks and Mitigation Strategies While we are feeling more confident with the optimization algorithm and the SUMO simulation platform, our camera situation is still uncertain due to an inability to connect to CMU-SECURE. We’ll be ordering a BLE camera to test our vehicle detection …
Work Done
This week, I prepared for and gave the design review presentation for my group. I also made some progress on the car detection code, but I realized that we will probably need to train our own Haar cascade, since the ones I found online don’t perform very well on settings where other buildings are in frame. In the screenshot below it’s evident that the classifier is fairly accurate:
But in this one, the classifier struggles to differentiate the buildings from the cars.
We’ll obtain some videos from local intersections with the camera angles that we’ll actually be using and test the pre-trained classifier on those; if we get similar lackluster performance, I’ll start training a separate one.
I also started the Raspberry Pi and IP camera setup; the IP camera setup is fairly intuitive as it uses an app interface to connect to WiFi (I haven’t yet attempted to connect it to CMU-SECURE, but got it to connect to my apartment’s WiFi.) The Raspberry Pi that we borrowed from the capstone inventory seems to have already been registered to CMU-DEVICE under an unknown hostname, so I’ve contacted IT to help us get that resolved (since we can’t ssh into it if we don’t have the hostname.)
Schedule
Because both the vehicle classifier and the Raspberry Pi setup are taking longer than usual, I am behind schedule as these should have been done by this week. I will keep working on the vehicle classifier through next week and spring break, and we may reduce our accuracy metrics for both the vehicle and pedestrian object counts if we continue to see issues with the detection code and repurpose a good portion of the vehicle detection code for the pedestrian detection code (since that’s what’s next on my schedule.)
Deliverables
By the end of next week:
Work Done I finished working on functions to call the APIs and set everything up for both TomTom and HERE. I also set up the SUMO simulation software on my laptop. This took really long because I was originally using my Macbook to install it, …
This week we were still in a very preliminary phase of researching/purchasing/planning before we can begin work on our implementation. We settled on the IP camera to order, ordered it, and reserved an RPi. I also did some research on Addressable LED Arduino projects and …
Work Done
This week, I contributed to the design review presentation with the rest of my group members (the hardware implementation plan, testing approaches, and system specification/block diagram.) I also tried to set up the Raspberry Pi and IP camera (unfortunately, we’re waiting on the microSD card for the RPi to come in so I wasn’t able to make too much progress with that.) I also made a first pass at the object detection code using a pre-trained Haar classifier for vehicles I found online. A screenshot of the code is below; the classifier is from this GitHub repo.
Schedule
I’m somewhat behind schedule, unfortunately, since I haven’t been able to connect the IP camera to the RPi or figure out how to connect them to CMU-SECURE. Once the microSD cards come in I should be able to get on that next week. I’m on schedule for the object detection code, since I’m supposed to finish it by the end of next week.
Deliverables
By next Monday, I will do the following:
Potential Risks and Mitigation Strategies At this time we feel more confident in our solution and we were able to finalize most of our solution approach with specific hardware and software we are using. Earlier in the week we were hesitant on how we can …
Work Done This week I worked on the Design Review presentation as well as doing additional research on our solution design, specifically on the APIs we will be using as well as the optimization algorithm. I have finalized the APIs we will be using to …
The most important task I accomplished this week was giving my team’s proposal presentation to Section D on Monday. Last weekend, Ankita, Kaitlyn, and I put a lot of effort into making sure we covered all of the necessary components for the proposal. This process helped us all to get a better idea of all of the details we would need to consider going forward in order to succeed with our project. Once we were all on the same page and had established a clear vision to all work towards, we put together our slides and decided as a group that I would be in charge of presenting this one. I ran through the slides to practice what I would need to say several times before Monday morning. I felt prepared going into it, but then ended up getting a little bit nervous at presentation time which caused me to trip up a little bit on the first couple of slides. I do feel like I recovered well, though, and regained my confidence throughout the presentation.
The rest of the week for me was dedicated to collecting traffic video data that we can use to train our CV model. I have managed to find some good resources online to pull footage from. One particularly helpful website is trafficcamarcive.com, which allows you to view live camera footage from many large cities in the country. There are hundreds of cameras available in Pittsburgh alone, so I don’t think quantity of footage sources will be an issue. The biggest challenge, though, I have faced so far with this task is making sure that the recordings are high enough quality and that the camera angle is oriented in a similar way to what we had envisioned in our ultimate implementation. I am also trying to find an array of different traffic situations and sort the videos into corresponding categories so that we can ensure we are training our models with a wide variety of data.
The plan for next week is to work with Ankita to figure out exactly how we will want to implement cameras in our product and then decide what cameras make the most sense to purchase based on factors like size, convenience, and cost. We will then get these ordered so that I can begin work on connecting them to our Raspberry Pi and testing that everything works as intended.
Potential Risks and Mitigation Strategies The main risk we currently foresee is being unable to get the IP cameras set up at an actual intersection to send data to the Raspberry Pi. We have a plan in mind for this (detailed in Ankita’s Status Report), …