本論文針對視覺里程計(Visual Odometry, VO) 運用幾個演算法對即時性進行改善。首先利用雙眼視覺系所統擷取影像，影像使用直立式加速強健特徵(Upright Speeded-Up Robust Features, U-SURF)偵測地標點。加入循序跳越式盒子濾波器(Ordinal Skip Box Filter)以減少在U-SURF中對影像迴積的次數與計算時間。之後利用參數化三點透視(Parameterized Perspective-Three-Point algorithms)演算法反推視覺系統可能位置。再者利用隨機化隨機取樣一致(Randomized Random Sample Consensus, R-RANSAC)演算法以剔除其他錯誤位置解。最後以地面基準實驗證明改善效果。 In this paper, several algorithms are applied to improve the instantaneity of the visual odometry system. First, the binocular stereo camera is used to capture images. The upright speeded-up robust features (U-SURF) algorithm is used to detect features in the images. Ordinal skip box filter added in U-SURF can reduce convolution times and computation time. Then the possible positions of the camera can be determined by the parameterized perspective-three-point algorithm (Parameterized P3P). Next, the randomized random sample consensus (R-RANSAC) algorithm is used to eliminate other error position solutions. Finally, improvement performance is shown on the ground truth experiment.
