- 學位論文
基於統計濾波器之點雲色彩去雜訊演算法超大型積體電路設計
VLSI Design of Point Cloud Color Denoising Algorithm Based on Statistical Filter
摘要
本論文提出了一種基於統計濾波的創新點雲顏色去雜訊演算法，目的在於提高點雲數據的品質，以在各種高階應用中得到更好的性能。點雲技術在室內設計、醫療研究、農業測量、環境評估、工業設計、航空航太和遊戲開發等領域的應用日益增加。相較於傳統的二維影像，點雲數據的精確三維建模能提供更豐富且直觀的資訊。點雲數據通過 LIDAR配合光學或 RGBD相機生成，然而在數據採集過程中，由於環境因素、相機缺陷、硬體干擾以及重建算法的錯誤，往往會引入雜訊。 本研究實施了一種低複雜度且高效率的演算法，著重在提升初始點雲數據的質量。該演算法的去雜訊過程包括全局濾波和局部濾波兩個階段：全局濾波針對整體點雲進行處理，以抑制微弱的雜訊並保留邊緣和特徵細節；局部濾波則 進一步細化平滑區域的去噪效果。所提出的演算法在 Greyc資料集中的去雜訊效果，相較於 spectral graph wavelet transform (SGW)方法，平均的峰值訊噪比 (PSNR)提高了 1.13db。 本研究進一步採用VLSI實現加速效果。硬體架構設計的目的在於加速運算速度。在提高速度的同時，還要保持高還原品質、降低功耗和壓縮面積。 在硬體設計上，採用 pipeline設計架構。由於點雲數據龐大，處理一張點雲圖通常需要進行數萬次運算，因此 pipeline架構能有效減少整體運算時間。 晶片設計使用 TSMC 180nm和 90nm製程 技術 。初期中值濾波器設計使用 TSMC 180nm製程，達到 100 MHz的工作頻率、9.76 mW的核心功耗以及 18,728的邏輯閘數量。改良後的局部濾波演算法使用 TSMC 90nm製程，工作頻率提高至 125 MHz，核心功耗降低至2.55 mW，邏輯閘數量為 67,514。 本論文所提出的演算法及其硬體實現提供了一個有效的解決方案，所提出的演算法及其硬體實現提供了一個有效的解決方案，用於提高點雲數據的質量，這對於高階應用至關重要。未來研究可以結用於提高點雲數據的質量，這對於高階應用至關重要。未來研究可以結合位置和顏色關係，以達成更精確的點雲圖像還原，並進一步優化硬體合位置和顏色關係，以達成更精確的點雲圖像還原，並進一步優化硬體架構以減少邏輯閘的使架構以減少邏輯閘的使用量。用量。
並列摘要
This thesis proposes an innovative point cloud color denoising algorithm based on statistical filtering, aimed at improving the quality of point cloud data for enhanced performance in various advanced applications. The applications of point cloud technology are rapidly expanding across fields such as interior design, medical research, agricultural surveying, environmental assessment, industrial design, aerospace, and game development. Compared to traditional 2D images, the precise 3D modeling provided by point cloud data offers richer and more intuitive information. Point cloud data is generated using LIDAR in combination with optical or RGBD cameras. However, during the data acquisition process, noise is often introduced due to environmental factors, camera defects, hardware interference, and reconstruction algorithm errors. This research implements a low-complexity, high-efficiency algorithm focused on enhancing the quality of initial point cloud data. The denoising process of the proposed algorithm includes two stages: global filtering and local filtering. Global filtering processes the entire point cloud to suppress weak noise while preserving edges and feature details, whereas local filtering further refines the denoising effect in smooth regions. The proposed algorithm shows an improvement in denoising performance on the Greyc dataset, with an average PSNR increase of 1.13 dB compared to the spectral graph wavelet transform (SGW) method. Additionally, this research employs VLSI implementation to accelerate the algorithm. The hardware architecture design aims to speed up computation while maintaining high restoration quality, reducing power consumption, and minimizing chip area. In hardware design, a pipeline architecture is adopted. Due to the large size of point cloud data, processing a single point cloud image typically requires tens of thousands of operations, so the pipeline architecture can effectively reduce overall computation time. The chip design uses TSMC 180nm and 90nm process technologies. The initial median filter design, using TSMC 180nm process, achieved a working frequency of 100 MHz, core power consumption of 9.76 mW, and 18,728 logic gates. The improved local filtering algorithm, using TSMC 90nm process, increased the working frequency to 125 MHz, reduced core power consumption to 2.55 mW, and used 67,514 logic gates. The proposed algorithm and its hardware implementation provide an effective solution for enhancing the quality of point cloud data, which is crucial for advanced applications. Future research could integrate positional and color relationships for more accurate point cloud image restoration and further optimize the hardware architecture to reduce the number of logic gates used.