Consistent and competent awareness in dynamic environment, specifically in solidly encumbered environments is still challenge nowadays. Most of the systems applied for target tracking are based on object models. Such strategies usually fail in complex environments with wide range of moving objectives. The tracking approach which is called Bayesian Model of prediction basically involves customary frames, where each cell stores information on the velocity distribution and grid occupancy. The cells are supposed to be absolutely mutually independent in order to avoid combinatorial explosion. The evaluation of cell tenancy and cell velocity is performed in a way which is similar to usual filtering, which includes the stages of prediction and evaluation. Sensory data from various resources (sensors) can be combined by the means of the approaches, which prediction model uses, as the cells are independent from each other, the perception of physical objects does not subsist in this space. It helps to eliminate the problem of data association which has to be determined in object based tracking systems.
The five principles of the Bayesian Model of Prediction acting are the following:
- Taking uncertainty into account explicitly. It is the general practice in such models and approaches, that uncertainty in the gesture recognition is inherent, and is usually noticed as a variant of occupancy frames.
- Avoiding the problems with data association. It means that the data association should be handled at the upper level of structuring. The approach is meant to use the following principles: the data association matters are usually meant to associate any object (Ot) at a time moment t with Ot+1 at time moment t+1. The described approach for solving the data association matters usually can not solve all of them properly in lots of circumstances. The concept of objects can not be regarded as existing in the Bayesian Model, and thus evades the problem of data association in the general tracking sense.
- Eliminating the object model problem. This point of the Bayesian Model strategy is meant to recognize the color, the shape and the magnitude of the object clearly. Definitely, it is really difficult to define what the sensor needs to measure and how it should define the object, and which object it should define from the variety of objects. The fact is that, some big object may affect various cells of the sensor, and cause unpredictable reaction of the system, while small object may affect only one cell, which may not be reacted by the system. Both cases may be caused by the lack of logics and coherence, which makes the most of the objective tracking systems work with the only object, and these systems are absolutely unable to define multiple objects.
- The Bayesian Model of Predictionh has been elaborated to be able to react on the multiple objects (i.e. to be parallelizable).
The fact is that, system may face the problems of changing of the appearance of the object because of the changes in lighting, angle variations, or image occlusions by the optics. This variability entangles the process of tracking. The tracking model should be able adapt to changes in object look over time.
The value of each pixel is created by a Bayesian combination allotment. A probabilistic model of a prediction is based on the expectation maximization algorithm and is regarded to be designed to enhance the parameters of model form and look. In creating the tracking algorithm, the probabilistic model is grounded on the adaptive appearance sculpt, and an atom filter is engaged in order to eliminate the unclear gestures or objects. The unauthorized pixels and occlusions are managed by the means of a robust-statistics method. Plentiful experimental outcomes show that the suggested algorithm can track objects well even with the changes of the lighting, large angle variations, and incomplete or full occlusions.
The designers of the Bayesian model of prediction in the tracking systems used about 3000 training illustrations from a variety of human activities including running, walking, conversation, quarrelling, pantomime, obtained using motion capture data. Computer graphics human model were used to generate guidance data including 54-dimensional joint angle condition vectors parallelized with matching, practically provided, image silhouettes.
Our described approach is discriminative rather than generative. It uses a huge human motion capture data-base and a 3D computer graphics human model, to generate training pairs of typical human arrangements, alongside with their practically rendered 2D silhouettes.
References
Sminchisescu, C. _ Kanaujia, A. (2004) “Learning to Reconstruct 3D Human Motion from Bayesian Mixtures of Experts. A Probabilistic Discriminative Approach” Technical Report, University of Toronto,