Research And Its Applications Of Computing Mean Shift In Real-time For Embedded Systems

With the development of Visual Things, visual processing algorithms will be increasinglyused in embedded systems. In practical applications, real-time calculation becomes a majorfactor to determine whether Mean Shift algorithm is practical. Therefore, how to improvecomputation speed of Mean Shift algorithm becomes a research hotspot. For Mean Shiftalgorithm, many researchers have done a deep research, but these algorithms are verycomplex, computationally intensive, and generally are implemented on PC or DSP. In manypractical applications, Mean Shift algorithm is difficult to meet the reality system forreal-time requirements. Therefore, to improve the existing algorithm, to search for moreeffective hardware implementation platform for solving the real-time problem in Mean Shiftalgorithm, it is a valuable subject for research.This paper firstly introduces the principle of Mean Shift algorithm in detail, and it alsointroduces parametric density estimation, non-parametric density estimation and Mean Shiftalgorithm. Then, it analyses the convergence of Mean Shift algorithm. Secondly, it introducesthe applications of Mean Shift algorithm in image processing. Finally, it introduces themethods of accelerating Mean Shift algorithm.This paper describes the application of Mean Shift algorithm in target tracking detailedly.First of all, it describes Mean Shift tracking algorithm in detail, and it also provides thespecific steps of algorithm. Then, it analyses the time complexity of Mean Shift trackingalgorithm. Finally, it introduces the factors of affecting the performance of Mean Shifttracking algorithm.This paper designs the hardware structure of Mean Shift tracking algorithm, and it alsorealizes it on FPGA hardware platform. First of all, it optimizes Mean Shift tracking algorithmin consideration of the characteristics of FPGA hardware platform. Secondly, it provides thehardware structure of the optimized Mean Shift tracking algorithm. It also describes how toimplement each module using hardware. Finally, it shows experimental results usingChipScope, and it also analysis the accuracy and real-time performance of experimentalresults. The system can handle78pictures per second, which size is288×384, and fully meetthe requirements of real-time process.