Copyright © [2023] Tao He, Marshal Guo.
A computer vision program that differentiates cars and bikes from the traffic in the video file, using only OpenCV and NumPy for image processing. Different training algorithms are provided in the comments of the train.py file for you to try out.
As this model is still a prototype. In a sample of 28 incoming traffics, only 4 vehicles were detected as the wrong category. Thus for this specific sample of testing data, the model has an accuracy of 86%.
In summary:
- Tidy the data
- read video and cancel background from moving objects
- select the targeted moving objects
- Train
- Run and Test!
Used the KNN segmentation background cancellation object provided by OpenCV and converted the color space of the frame to YCrCb. Then applied the BackgroundSubtractorKNN object to the frame, obtaining a binary mask where moving objects are highlighted. Also applied median filtering to the binary mask to reduce noise.
Moving objects are marked white and static is black
Find each highlighted moving object that is bigger than 100 pixels and export the images as raw data.
Looks like the correct objects are selected!!
To train the machine, we need each image data to have the same amount of pixels. Thus, we find the largest image in the raw data and use its dimension to enlarge the rest of the data to have the identical dimension.
Then according to the different width-length ratio of the "white objects", we put cars and bikes in separate datasets in the train_data folder.
Two resized bike and car images
Iterating over the subdirectories of train_data and collecting training data and labels, program converts them to grayscale using cv2.cvtColor(), and then flattens each grayscale image into a 1D array (x_img1) using gray1.flatten(). The flattened image array is appended to the list xtrain, and the corresponding label (subdirectory name) is appended to the list ytrain.
Result is as follows!
End Result Voila!