**6. Acknowledgment**

288 Real-Time Systems, Architecture, Scheduling, and Application

This chapter has approached the problem of the real-time implementation of motion estimation in embedded systems. It has been descripted representative techniques and systems belonging to different families that contribute with solutions and approximations to a question that remains still open, due the ill-posed nature of the motion constraint

It has been also proposed an overview of different implementations capable of computing real-time motion estimation in embedded systems delivering just low primitives: (gradient model and block matching respectively) and delivering the mid-level vision primitives (combination optical flow with orthogonal variant moments). In the Table 7 are shown the different methods implemented regarding the machine vision domain, the final performance obtained, the robustness implementation and the complexity of the final

These systems designed are scalable and modular, also being possible choice the visual primitives involved -number of moments- as well as the bit-width of the filters and computations in the low-level vision -optical flow-. This architecture can process concurrently different visual processing channels, so the system described opens the door to

The implementation of these systems shown offers robustness and real-time performance to applications in which the luminance varies significantly and noisy environments, as

> Throughput FPS (320x200)

Moments Low&Mid 59 fps Low Easy

(McGM) Low 26 fps Very High High

Search Tecnique) Low 120 fps Medium Easy

Sensor. Low&Mid 26 fps Very High Very High

Robustness Complexity of the

Implementation (FPGA board)

industrial environments, sport complex, animal and robotic tracking among others.

the implementation of complex bioinspired algorithms on-chip.

Table 7. Comparison of the different methods implemented.

Vision Domain

Table 6. Throughput in terms of Kpps and fps for the block matching technique.

Throughput 12 Mpixels/sec 20 Mpixels/sec 30 Mpixels/sec

Quality I Quality II Quality III

COMPLETE Mid-level and Lowlevel Vision

**5. Conclusion** 

equation.

system.

Method Implemented

Orthogonal Variant

Motion Estimation

Multimodal Bioinspired

Block Matching (Full

This work is partially supported by Spanish Project project TIN 2008-00508/MIC. The authors want to thank to Prof. Alan Johnston and Dr. Jason Dale from University College London, (UK) for their great help with Multichannel Gradient Model (McGM), explained partially in this chapter.
