This project is do identify cars parked in a certain parking spot. To achieve this goal, the main steps are as following:

• Download the video files and extract the first frame as input image

• Use YOLOv3 to detect objects in an image

• Use SIFT (Scale-Invariant Feature Transform) algorithm to find keypoints and descriptors from an image, then we can match local features to each other and therefore compare images to each other

• Python 3.6
• opencv 3.4.2 (lastest version doesn't support SIFT detector)
• yolov3.weights
• coco.names

Use OpenCV bulid-in function VideoCapture() to read the video, save the first frame in a file named image

• I didn't process all the video files. I choose the 2000th video as start point and sample every 60 videos to 8000th and get the first frame.

• Detection Using YOLOv3 Model

  • Install Darknet

git clone https://github.com/pjreddie/darknet
cd darknet
make

- Download the pre-trained YOLO v3 weights file

wget https://pjreddie.com/media/files/yolov3.weights

• Extract the detected car or truck in the certain spot

• find the keypoints and descriptors with SIFT
• FLANN based matcher to match the keypoints found in last step
  • FLANN stands for Fast Library for Approximate Nearest Neighbors. It contains a collection of algorithms optimized for fast nearest neighbor search in large datasets and for high dimensional features.