| 摘要: | 為了清楚地了解合成孔徑雷達(Synthetic Aperture Radar, SAR)複雜的特性，特別是當近代的雷達系統逐漸走向高解析、多波段、全偏極 和混和多種觀測模式的系統下，需要萃取出更多目標物特徵來輔助判斷。此篇文章中的SAR影像資料庫模擬器能發揮以下多種功能。為了達到輔助目標物辨識的目的，多重角度的特性囊括於此模擬器中，利用此多重角度的資料庫以降低SAR目標影像的角度高靈敏問題，藉此提高辨識的精確度。此外此模擬器亦可用在SAR散射機制的了解上，藉由分離不同反射和尖端散射下的背向散射結果圖，得而清楚地了解何處的散射電場分布是肇因於何處複雜目標的結構上。透過這樣的各層次的背向散射結果圖，特別適合用在複雜目標模型上，得以區別哪些是因目標或背景的影響。在計算機輔助工程(Computer-aided engineering, CAE)的目的上，能透過前期的軟體模擬適當地降低實驗的風險，因此無論是在SAR處理中的演算法亦或是硬體階段的開發，皆可仰賴此模擬器的優勢。其他例如硬體模擬器的應用，亦可使用此模擬器提供後端的軟體資料產製，這樣的硬體模擬器主要是透過軟硬體整合以提供複雜的人造SAR訊號，再由硬體發射。
本篇文章中提供了完整的模擬器流程，由模擬參數表的產製至完整的多重角度SAR目標資料庫。有別於相關的文獻，此模擬器結合目標物與背景的交互關係，並著重在多重角度的SAR目標物影像產製，從而有相應的平行化處理需求與實作，亦使此模擬器更具真實和可行性。在實作此模擬器上運用了諸多技巧。在背向散射點場的計算上，同時考量了多次反射與尖端散射效應，相應而生的光線追跡法得以滿足多次反射的需求，物理光學法(Physical Optics, PO)以近似解表面電流，物理散射法(Physical Theory of Diffraction, PTD)得以求解楔型尖端散射。多層與不同材質的表面亦考慮於其中。為了求取高精度表面積分問題，文章中採用了泰勒展開式逼近法。綜合以上方法，可使背向散射電場不僅局限於單一反射面下的散射現象，而物理散射法特別適合利用在目標物高複雜度的邊緣結構。在斜距向的取樣點上，為減少計算複雜，所有計算上皆為頻率域的樣本點，藉此頻率樣本點數可限縮在感興趣的斜距跨幅內，實為相依於目標物的大小，而獨立於目標物的斜距距離。為求更真實，基於星載和機載的感測器軌跡模擬亦包含於此模擬器中，諸如感測器量測裝置上的偏移，抑或雜訊誤差等皆包含於其中。更重要的為，考慮多種SAR觀測模式，包含Circle SAR、Inverse SAR、Stripmap和Spotlight等模式，可將大部分現行運作中的SAR系統囊括於此模擬器中，相應而生的成像方法諸如Range-Doppler Algorithm (RDA)、Omega-K Algorithm (WKA) 和 Time-domain Back Proejction (TDBP)等，都需搭配在不一樣的觀測模式下。
綜觀整個模擬器，由於此計算量即便在單一目標下實為龐大，則在資料庫的產製上更為不可能，因此基於NVidia CUDA技術則用來加速整體模擬器。根據程式碼剖析，在背向散射電場計算是為主要最佳化平行的區塊，經過層層的加速測試，最終最佳化的平行化程式可將速度提昇至少200倍以上，而複雜目標物的組合三角片可多達百萬片，在大區域目標上經實測可達到250公尺 x 250 公尺的範圍，而單一角度下的影像可達到近乎30分鐘的計算時間，因此推論此模擬器即便在非大型工作站下及有其模擬的實用性。最後使用的MSTAR量測車輛資料模擬小範圍高解析度的複雜目標物影像，而英國白金漢宮模型資料則用於實現大範圍的場景目標，其中目標物的尺寸維度與散射特徵亦包含於測試結果中。;In order to understand clearly the complicated characteristics of synthetic aperture radar (SAR) especially in high-resolution condition, multi-band, fully polarization and multi-acquisition mode, more and more features need to be extracted. For recognition or identification purpose, the multi-aspect angle database is essential because of the accuracy of classification. The high angle sensitivity of SAR target image was considered. For the SAR mechanism understanding, a distinction of the backscatter electric field between each reflection and diffraction is necessary. According to the separated scattering electric field map, which reflections are due to which structures of complex target can be clearly identified. This simulator is very suitable to understand these characteristics. For the computer-aided engineering (CAE) purpose that is adopted for the engineering analysis to avoid the risk of experimentation widely, computer software orientation simulation is the earliest motivation. For the evaluation of SAR algorithm or system this simulator can be adopted to be the first version before lunching or flight-testing. For the hardware orientation emulator application, a combination of software and hardware is very suitable for providing the man-made artificial SAR complex target’s signal.
An overall simulation flowchart from simulation parameter tables to SAR target’s database was introduced in this dissertation. Different from the literature are that the interaction between target and background was considered, the purpose of this simulator was multi-angle orientation, and the optimal parallelization architecture was introduced by power of CUDA that make this simulator be authentic and usable. To implement this simulator, a lot of technologies were considered. The simulation of the back scattering electric field must consider the multi-bouncing back scattering and diffraction. For the purpose of parallelization, the ray-tracing technology was adopted. After the ray tracing, the Physical Optics (PO) and Physical Theory of Diffraction (PTD) methods with multi-layer material was adopted. For more high accuracy consideration, the Taylor expansion method was used for solving the integration of surface current. Therefore, the back scattering electric filed is not restricted in the single bounce. Furthermore, an analytical diffraction solution was adopted for the wedge structure that is very common on complex targets in CAD model especially. For reducing the computation time, a slant range sample in frequency domain was adopted to limited the number of samples in the swath region. Furthermore, the sensor path trajectory for space- and air-borne was considered. For the authenticity, bias and noise can be generated in this simulator to meet the real sensor requirement. Because that this simulator is for producing appropriate SAR databases, the SAR geometry combination with different kinds of observation type was considered such as the Circle SAR, Inverse SAR, Stripmap and Spotlight. Various SAR focusing algorithms such as the Range-Doppler Algorithm (RDA), the Omega-K Algorithm (WKA) and the Time-domain Back-Projection (TDBP) were considered too.
Base on this simulator, the heavy computation loading was profiled and figured out by a flowchart. The back scattering computation block takes a high complexity with order of cubic. The NVIDIA CUDA technology was adopted, a high speedup rate was represented in more than 200 times in the experimental dihedral results. It means that this simulator is very suitable for very complex target where number of polygon is more than one million, and for larger target that the dimension is larger than 250 by 250 meters. The computing elapsed time is almost a half hour per SAR image clip. This calculation speed makes this simulator be feasible. Finally, an MSTAR (Moving and Stationary Target Acquisition and Recognition) vehicle and larger Buckingham palace complex target models were applied to the simulator in the final chapter. The verification of target’s dimension with feature can be found those results. |
