Human motion analysis is currently one of the most popular research fields in computer vision. It deals with the detection, recognition and tracking of people. It's a good way to understand human behavior by processing of image sequences involving humans.
In our project, facial recognition system is for identifying a person from a video frame from a video source, through the way of comparing selected facial features. Video tracking is the process of locating multiple persons and their faces simultaneously in a video sequence by analyzing the video frames. The tracking outputs the location of moving targets within the video frame.
Our goal is to recognize and track multiple persons in a video clip. To simplify this problem, we track the face rather the whole body of each person who shows up in the video.
1)As the first step, we need to detect faces from an image,which is usually the first frame of the video, and find out the locations of those faces. These locations can be used as the initial input values for tracking.
2)Once we are able to get the initial positions, we want to track persons who show up in the video-that is, to locate the face of every single person in each frame of the video. It is possible that a person's face color is similar to the background, or more than two people get closed to each other and their faces overlap int he video. Both situations will make it difficult to track our objects. Addressing this issue, which is known as the "occlusion problem", is an significant point in our project.
1)For the first step, we plan to use Adaboost algorithm to detect faces. We might use some human face photos as data set to train the detector.
2)For the second step, We will use Condensation algorithm (Conditional Density Propagation) to track the multiple person. Probabilistic algorithms are essential for robust tracking in the presence of dense background clutter. While Kalman filter has been limited in the range of probability distribution of probability. Condensation allows quite general representations of probability which is more appropriate for our multiple-person tracking goal.
We will take a video record indoor as our dataset. Basically, two or more people move in the video. Their faces may completely or partially overlap each other.