MPEG-4/H.264視訊壓縮標準於嵌入式系統之最佳化研究

Abstract

本論文中提出許多最佳化之方式將MPEG-4與H.264視訊壓縮技術分別實現在以ARM、FPGA搭配之SoC開發平台與以ARM、DSP為基礎之雙核心平台，透過硬體、韌體及軟體上在壓縮流程上的最佳化，將複雜的演算法實現在運算能力較低之嵌入式系統中，並得到良好的壓縮效果。 在MPEG-4硬體最佳化上，本論文提出一種階層式的動態估測演算法，透過數學的分析以及實驗的結果來決定哪些步驟對於壓縮品質之影響力較為廣泛，推導一快速、高品質且適合用硬體實現之演算法。且其運算複雜度僅為全域搜尋法之3.91%。搭配此演算法，以ARM配合FPGA的實現方式來進行MPEG-4即時壓縮系統的研究。此系統可依據使用者要求動態調整輸入以及輸出影像參數，可有效提高其應用性。 而在H.264於ARM/DSP雙核心嵌入式平台的實現上，推導一適合影像式車用防盜系統之動態估測演算法，可用低複雜度的運算方式，有效克服車輛行駛中因外在光源變化等因素導致視訊壓縮品質下降的問題，並且針對車輛影像特性將動態估測流程作動態更新，能在重點區域保持良好壓縮品質，同時大幅增加運算速度。除此之外，本論文亦將H.264壓縮流程做最簡化，利用記憶體配置管理、雙核心CPU工作分配以及DSP特有指令等最佳化技巧，成功移植入低成本且高穩定之嵌入式系統中。 除了將演算法以硬體與軟體等不同的方式進行最佳化之外，本論文亦將實現之系統實際運用於由交通大學自製研發之Taiwan iTS-1智慧車中，搭配了Zigbee無線傳輸以及3.5G/3G手機程式設計，當車輛正在移動時，本系統可偵測駕駛員前方、後方以及兩側盲點區域之車道線以及車輛距離，並以直覺的方式提供資訊。而當車輛停妥後，若有歹徒強行闖入，系統則會即時發送警告訊息至使用者，車主即可利用手機隨時隨地瀏覽愛車情況，並可馬上提供警方歹徒之面貌。本論文之宗旨在利用各種最佳化方式，於嵌入式系統上實現動畫壓縮技術，並將其與智慧型車輛結合，讓駕駛員、乘客以及車輛都能有更安全的環境。

In this dissertation, several optimization approaches for implementing the MPEG-4 visual and the H.264 video processing standards on an ARM/FPGA Soc platform and an ARM/DSP dual-core embedded system are addressed. By the modification of the compression procedures in hardware, firmware and software, the complex algorithms are successfully ported to the low computing power embedded devices, and good encoding performance is obtained. In the MPEG-4 hardware optimization, an efficient hierarchical motion estimation algorithm (HMEA) is proposed. With mathematical analyses and the experimental results, the search steps, which can dominate the image quality, are designed in order to develop a high-speed and good quality architecture that is suitable for hardware implementation. The operational complexity of HMEA is only 3.91% of the full search block matching algorithm, and with HMEA, a register-based platform-independent MPEG-4 co-processor (RPIMC) is designed. RPIMC can change the input and the output parameters of the bitstreams by adjusting the value of the registers dynamically to widen its applications. For implementing the H.264 framework on an ARM/DSP dual-core embedded system, an adaptive motion estimation algorithm (AMEA) especially for the vehicle surveillance videos is developed. By easy manipulations, AMEA can overcome the coding quality degradation due to the changing environments, and it can update the search procedure adaptively according to the video characteristics in a moving object to maintain the performance in the region of interest. Moreover, the platform-based acceleration techniques, such as the memory organization, the dual-core communication and the program modification, for H.264 are also presented. In addition to algorithm optimization in hardware, firmware, and software, the developed system in this dissertation has been successfully tested in the intelligent vehicle, Taiwan iTS-1, developed in National Chiao Tung University, on the real road environment in Taiwan. Cooperating with the warning system, which can detect the vehicles and the lane marks in front/rear directions and blind spot areas, the Zigbee wireless communication and the 3.5G/3G mobile phone programming, the system can provide the distance information of the surrounding cars to the drivers through a friendly intuitive graph, and it will produce the warning signal when lane departure. After parking, if there exists intruders, the system will send a notification message to the users, and they can browse the real-time images of the vehicles by their mobile phones anytime and anywhere. The main contribution of dissertation is to provide a safer environment to drivers, passengers and vehicles at all time using several optimization approaches for implementing complex video processing standards on embedded systems.