**2.4 Configurable hardware algorithms implemented for optical flow**

There are several real-time hardware systems founded on the algorithms mentioned herein, proceeding to do a quick review.

• Some algorithms used are of matching or gradient, such as Horn and Schunck algorithm (Horn & Schunk, 1981) that has been carried out using a FPGA (Zuloaga *et al*., 1998; Martin *et al*., 2005). The model used is straightforward, is not robust and does 276 Real-Time Systems, Architecture, Scheduling, and Application

The schemes discussed so far, consider the calculation of optical flow as a separate problem for each frame, without any feedback (the results of motion of a frame does not cover the analysis of the following). Giaccone and Jones (Giaccone & Jones; 1998, 1999) have designed an architecture capable of dealing with multiple motions (keeping the temporal consistency)

This algorithm has proven to be robust for a given speed limits, also works well when compared to similar models. The cost calculation is overwhelmed by the generation of a projected image, PAL sizes needed for about 40 seconds/image, with SPARC 4/670MP.

However, this time consistency constraint is only used sporadically today. The objects in the real world must obey physical laws of motion and the inertia and gravity, so that there is predictability in their behavior, and at least surprising, that most real-time algorithms do

For this purpose, it is used the fact that it is possible to create an additional constraint equation from the velocity field for use in the next iteration, managing the problem as an evolutionary phenomenon. The use of probabilistic models or Bayesian (Simoncelli & Heeger, 1991) may be an alternative to using real world information and update results of

We have seen that the perception of motion can be modeled as an orientation in space-time, where the methods of extracting this orientation gradient across the filter ratio oriented. The so-called motion energy models are often based or similar in many respects to models of gradient, since both systems use filter banks to obtain this time-space orientation, and therefore the motion. The main difference is that the filters used in energy models, are designed to meet time-space directions, rather than a ratio of filters. The design of space-

The methods of energy of motion are biologically plausible, but the implementations have an extra computer associated high due to the large number of filtering required being difficult its implementation in real-time. The resultant velocity of energy methods is not obtained explicitly, unlike gradient methods, only using a solution population, being these

One advantage, is that the bimodal velocity measurements as invisible movements, can be treated by these structures (Simoncelli & Heeger, 1991). The correct interpretation of the processed results is not an easy task when dealing with models of probabilistic nature. Interesting optimizations have been developed to increase the speed of these methods,

There are several real-time hardware systems founded on the algorithms mentioned herein,

• Some algorithms used are of matching or gradient, such as Horn and Schunck algorithm (Horn & Schunk, 1981) that has been carried out using a FPGA (Zuloaga *et al*., 1998; Martin *et al*., 2005). The model used is straightforward, is not robust and does

which segments moving regions by using a method of least squares.

not implement a flow based feedback.

last Bayesian models.

proceeding to do a quick review.

previous estimates integrating temporal information.

time oriented filters is usually performed in the frequency domain.

combined with Reichardt detectors (Franceschini *et al*, 1992) as support.

**2.4 Configurable hardware algorithms implemented for optical flow** 

not provide optimal overall results in software, but the implementation is efficient and the model is capable of operating in real time. The design uses a recursive implementation of the constriction of applying a smoothing iteration in each frame.

