| 摘要註: |
在近幾年三維技術的蓬勃發展下,特別是用於三維重建的儀器與裝置,因缺乏移動感測技術,導致資訊的取得不完整,這會使系統無法在較大範圍的區域獲得資訊的擷取,造成三維資訊來源的不足。本研究提出一方法以提高雙眼視覺系統之本體移動估測,能使裝置得知本身的移動狀態,得到更完整的資訊。首先影像擷取以The KITTI Vision Benchmark Suite提供之立體視覺資料,將經由校正過的左右相機對正影像,特徵點偵測部分以加速穩健特徵演算法(Speeded Up Robust Features, SURF)偵測特徵並產生敘述向量後,進行半全域匹配(Semi-Global Matching, SGM)以確保匹配之品質與數量,結合相機參數計算影像中之特徵點的三維座標。在對影像中特徵點位置進行時序性匹配,利用卡爾曼濾波器與適應性匹配的方式進行特徵追蹤,再從計算之三維座標、二維投影位置與相機參數以有效N點透視參數估測演算法(Efficient Perspective-n-Point Camera Pose Estimation, EPnP)得知相機的移動方式,最後採用稀疏光束調整法(Sparse Bundle Adjustment, SBA)取代BA (Bundle Adjustment)進行最佳化估測。實驗顯示在周遭移動物體不多的情況之下進行路徑估測,能達到平均每移動一公尺長度誤差在2公分以下,角度誤差在0.02以內的高精準結果,而在大多數的路徑情況長度誤差也能保持在每一公尺在5.5公分以下,角度誤差在0.05度的表現,最重要的部分是比起只使用適應性匹配方式在路徑複雜度較高的情況下會有不穩定的狀況,根據卡爾曼增強其適應性匹配,在彎道複雜度高路徑估測上有著更好的表現。 In recent years, many novel 3D technologies have been developed, particularly, systems and devices for the acquisition of 3D data. However, such devices often lack motion sensing components, causing difficulties in the registration of local 3D data in a larger environment, and attenuating the compatibility of 3D data with other sources. We propose a method to improve the ego-estimation calculated by means of binocular stereo vision, such that the 3D acquisition system is able to determine its own motion and use the information for the integration of 3D data. The input images are from The KITTI Vision Benchmark Suite, which are calibrated pairs of binocular stereo images. The feature points are detected using the Speeded Up Robust Features method, and descriptor vectors are generated for the features. Semi-Global Matching is then performed to ensure the quality and quantity of the matching features, which are used in conjunctions with camera parameters to calculate the 3D coordinates for the image features. Kalman filter and adaptive matching are the used to perform feature tracking. The calculated 3D coordinates, as well as the 2D projections and camera parameters are used in Efficient Perspective-n-Point Camera Pose Estimation to find the camera motions. Finally, Sparse Bundle Adjustment is used to perform optimization of the estimated motions. Experiments have shown that using the proposed method for motion estimation, the average translation error over 1 m is less than 2 cm, and the rotation error is within 0.02 degrees. In more complex scenarios, the proposed method is able to maintain the translation error to within 5.5cm and rotation error under 0.05 degrees. The most significant improvement is that the proposed method is able to provide more stability for paths with significant direction changes and demonstrate better performances with improved ego-motion estimations. |