Detailed query of TLD

The reason why the TLD algorithm is used is that it is classic enough, involves a wide range of topics, and references are classic enough. Haha, the original author is already preparing for a company. At the beginning, I started learning from it, gradually expanded it, wrote it down, and gave it to beginners for a little reference.

TLD code is open-source. You can select any object from the camera or test video for tracking. Therefore, TLD can be specific to something compared with other algorithms using classifier, take hog + SVM and ISM for example. These two algorithms are specific to a specific category, such as pedestrians, vehicles, this is mainly based on the classifier. during training, You must select training samples and sample tags, that is, supervised learning. Of course, you can learn online, that is, semi-supervised, but this requires a high real-time classifier. In incremental learning for robust visual tracking, Y uses online learning, however, I have simulated that efficiency is not enough for real-time applications. TLD also uses online learning, which is more efficient because it uses random.
Forest classifer is a good classifier and popular in recent years. foreign experts in the CV field I know are limited, but they use a lot of this RF. TLD also has an innovative place is P-N learning, then it is his side of detection and tracking. Maybe you still don't know. It doesn't matter. Let's take a look at it step by step. Let's take a look at the overall picture below:

This figure shows the TLD Algorithm Framework and explains tracking-Learning-detection. Its tracker uses the LK method. This is a frame.
The frame tracking method is used to estimate the movement of the target. Of course, lk tracking is not accurate and will fail to be tracked. Our RF detector will be used when the tracking fails, detector is used to locate the target of each frame and record the corresponding location information and surface information. It is called learning. When a tracker is lost, it relies on the past learning of detector, relocates the target. Learning is the link connecting tracker and detector. This learning method is P-nlearning.