多重應用自適應十字視窗雙鏡頭距離偵測技術與其系統實現

Abstract

本論文提出多重應用自適應十字視窗雙鏡頭距離偵測技術，具有低複雜度及高平行度的演算法，此演算法包含自適應十字視窗雙鏡頭距離偵測以及基於深度圖之侵蝕演算法。只需考慮周圍13個像素的顏色及紋路來產生深度圖。計算時間與原演算法相比達到2.97倍的加速。 整合自適應十字視窗雙鏡頭距離偵測與投票濾波器、中間值濾波器以及可消除近距離的雜訊的基於深度圖之侵蝕演算法，來產生一個對於Middlebury benchmark 只有15%誤差率。此演算法在160x120解析度下，可達到11fps之執行效能卻與原解析度的深度圖只有微量的誤差，並且可提供移動物件偵測。此外，藉由調整雙鏡頭的間距，本系統可以達到多種不同場景的應用。

This thesis proposes an adaptive cross-window stereo camera distance estimation technology which is a low complexity disparity algorithm and high efficient parallelized, which includes Adaptive Cross Window Distance Estimation (ACWDE) and disparity-based erosion. Calculating color and texture of13 pixels to generate disparity map. The computation time is speeded up 2.97 times compares to the original system. By integrate ACWDE with voting filter, median filter and proposed disparity-based erosion, which is erodes noise only for nearer disparity level, good enough disparity map can be generated, which 15% average error rate in Middlebury dataset. This proposed system can reach 11fps at 160x120 resolution on desktop with less difference disparity with VGA resolution. The proposed ACWDE algorithm also facilities moving object detection. Furthermore, this system can be adopted in various applications by changing the baseline of stereo camera to set the distance estimation range.