We introduce the basic concepts for detection and tracking of objects in videos. Several different moving object detection algorithms are presented that allow detection of objects. These objects are then tracked, using first a Kalman filter, which allows prediction of the future location of the target, and then the Condensation Tracking algorithm, which allows the tracking algorithm to keep multiple hypotheses of the object, and to model expected state changes.
We introduce some applications where target detection and tracking are needed. See some example videos:
We will need to have algorithms for robustly estimating the mean and standard deviation of a set of values, where there might be outliers in the set.
We introduce several moving target detection algorithms based on differences between frames or with a static background.
When the background is changing due to lighting or leave motion, etc, simple subtraction is inadequate.We introduce a non-parametric statistical approach to modeling the background and then detect foreground pixels that are unlikely to be part of the background.
We demonstrate the adaptive change detection algorithm on an indoor scene where the lighting is slowly changing, and where there are compression artifacts in the video.
We present several heuristics from removing small detections that arise because of noise, compression artifacts and camera jitter.
Once some potential targets have been detected in each image, they need to be linked together to describe the path that the detected object took, which can then be analyzed to help understand the behavior of the target. We assume an relatively sparse scene with no occlusions.
We introduce the Kalman Filter, which is an important algorithm for estimation of variables given data that arrives periodically. Here it is used to track the ball, predict its position and refine the position estimates.
The Kalman filter is applied to the falling ball tracking problem.
When a target can change state, tracking can lose the target. If you have a model of different possible successor states, then the Condensation tracking algorithm can explore these probabilistically. Also, it keeps multiple hypotheses so it can potentially have a better estimate of the state.
We extend the ball tracking example to use Condensation tracking, which allows tracking of the ball through bounces and stopping.
We consider the strengths, weaknesses and some extensions of the presented algorithms.
Automated video analysis systems have the potential for unethical use. The video introduces some potentially contentious applications.