The past week, I primarily was working on the design report, and since we had to pivot our design from counting people in Sorrells Library entrance to mapping a heatmap for Hamershlag 1300 wing hallway, we had to clear up a lot of details in our new design.
I got tracking working on a local computer using Deepsort with the webcam as input as well. After spring break, I will try to work with Gary and test the CV system with video footage of Hamershlag 1300 hallway (ideally sampled with our camera module). This will allow us to see how our tracking and detection system work within our target space. I will also start working on the interpretation of output based on our sample footage during the next week. Currently, the FPS of our CV system is around 5-6 fps, which is lower than our requirement. During testing with video footage of Hamershlag 1300 hallway, I will also try using faster detection models and faster tracking algorithms to compare the results and find a good balance between the accuracy of detection/tracking and speed. Currently, I am slightly behind schedule as I had to spend unexpected amounts of time on other classes in the previous week. Luckily I do not have many assignments in the upcoming week and will use that time to catch up.
New tools I have to learn to complete the backend subsystems I am responsible for are:
OpenCV:
- processing image and identify ROIs
- potentially explore tracking algorithms in openCV such as MedianFlow
Explore other detection algorithms to improve performance:
- Sparse Yolov5
- Yolov8
Create Decision Tree for Multi-class attributes:
- I only have experience working with decision trees as a binary classifier, will need to look into building a decision tree for 4 classes as output and multi-class inputs for our prediction system
Using SQLite with python:
- Need to learn using SQLite in conjunction with python to store backend data
