Methodology
First, we build the appearance model by searching for all the information needed to describe the target object:
- Video information: We retrieve the video and its technical information (number of frames, height, width, etc.)
- First frame information: We retrieve the information from the first frame to model the appearance of our tracked object:
- The frame: Very narrow frame to contain the tracked object and eliminate as much background as possible
- The color histogram: Color histogram to describe the information in the frame, therefore the object
The particle filter: We generate a random number for our filter based on the height and width of our frame and the number of particles we want to use.
Secondly, we build the motion model to predict its location in the next frames:
- Initialization: We initialize the histogram tracking.
- For each frame, not counting the first:
- Retrieve the frame and its color information
Activate the color histogram tracking that compares the frame with the previous frame (In my code independently unfortunately)
- Search for the same frame as in the previous state
- Generate new frames for the future frame
Then for each, check its similarity with the frame that is in the previous frame according to its state.
calculate the cumulative probabilities for each frame to determine the current state among them. It will be the one with the highest weight.
- Display the current state of each frame with a point.