"Mathable Score Calculator" and "Video Surveillance of On-Street Parking Spaces" were individual projects made during the Computer Vision course taken in the 1st year of the Artificial Intelligence master program at the Faculty of Mathematics and Computer Science, University of Bucharest.
The project consisted of calculating the scores of two players during multiple iterations of the board game Mathable. A multitude of images were given (regular, rotated, and perspective views), showing the players' placed tokens as well as the state of the game board during the players' moves.
Click here to see the datasets, as well as other project details.
The approach taken to realize this project involved the use of the following computer vision methods:
- template matching: in the beginning, templates of each token and type of board tile were extensively cropped, so as to recognize the newly placed token in each player's move
- SIFT features: in order to "straighten" rotated and perspective images of the board
- projective transformations: used alongside SIFT features
Note: for further details about the project solution, please consult this path inside the repository
This project solution achieved correct results on all of the given input data (see the link to the datasets from above).
The goal of this project was to solve four tasks tied to the surveillance of diagonal on-street parking spaces and two traffic lanes:
- Task 1: determine which parking spaces are occupied and which are not, from an image and a txt file of the given parking spaces
- Task 2: the same requirements as task 1, but this time checking all of the parking spaces instead of just a few given ones
- Task 3: tracking a specific car during a short video of up to one minute
- Task 4: identify how many cars are stopped at a red traffic light at the last frame of a similarly lengthy video to the ones at task 3
Click here to see the datasets, as well as other project details.
In order to solve this project, the following methods were employed:
- For all of the tasks, a pretrained version of the YOLOv9 model was used, namely the YOLOv9e detection model from the Ultralytics website
- Tasks 1 & 2: because the model had trouble detecting parking spaces that were farther away, a manual cropping of the last 5 spaces was made and then fed to the model
- Task 3: the tracking was made by calculating IoU scores at each frame between the last retained bounding box and all of the bounding boxes detected in the current frame
- Task 4: similarly to tasks 1 and 2, a multitude of croppings were made in order for the model to more accurately detect farther away cars, and only the last frame of the video was taken into account
Note: for further details about the project solution, please consult this path inside the repository