基於立體視覺之即時障礙物追蹤與避障方法

Real-time Obstacle Tracking and Collision Avoidance Methods Based on Stereo Vision

莊凱強(Kai-Chiang Chuang)

指導教授 ： 林達德

國立臺灣大學/生物資源暨農學院/生物產業機電工程學研究所/碩士(2013年)

https://doi.org/10.6342/NTU.2013.01780

摘要

本研究以兩顆CCD攝影機建構一套雙眼視覺立體影像系統藉以作為車輛與農業環境監控預警系統之應用。此系統將可計算出影像中像素的三維資訊，然後利用此資訊投影至上視圖進行障礙物偵測，再設定幾何限制將障礙物標定於畫面中。針對偵測到之移動障礙物影像，以巴氏距離 (Bhattacharyya distance) 進行結合距離與色彩特徵的計算，再匹配以達到追蹤目的，追蹤到目標物除了可計算取得運動的行進資訊與速度外，還可依此判斷該障礙物之動態行為作為警示作用。除此之外，追蹤到的障礙物其座標位置將會被紀錄並利用卡爾曼濾波器 (Kalman filter) 進行軌跡預測。此外，先透過立體視覺對前方資訊進行地圖的建立，再藉由A* 路徑規劃演算法達到導航與避障的作用。綜合立體視覺系統所得到的環境資訊，以抬頭顯示器之概念加以整合與顯示作為系統的避障模式。經過實驗驗證，本系統已能成功應用於車輛導航和農業的即時環境監測，透過提出的適應性巴氏距離和交叉比對方法在不同場景皆能達到95% 以上的追蹤準確率，軌跡預測誤差平均約25公分。在路徑規劃演算法A* 的應用下，進行區域式地圖搜尋，其搜尋範圍平均約49% ，經實驗驗證能有效應用於即時導航。

關鍵字

立體視覺 ； 巴氏距離 ； 卡爾曼濾波器 ； A*演算法

並列摘要

In this study, a dual camera stereo vision system was built to apply in the pre-crash warning system of vehicle and agriculture environment monitoring. Based on this stereo vision system, the three-dimensional information of each pixel in the image can be estimated. Then, this information was projected to top-view so as to detect and mark obstacles according to some geometric limited. For these detected moving obstacles, we combined depth and color features and calculate Bhattacharyya distance to match different obstacles between the two continuous images in the video sequence. While the targets have been tracked, the motion models will be built. Meanwhile, the state of obstacles and their speed can also be estimated which is useful for warning. In addition, Kalman filter was employed in location prediction with these models, and then we achieve navigation and obstacle avoidance by means of A * path-planning algorithm. Finally, the concept of head-up display design is applied to integrate with the above information ahead of vehicle, which can help users to make correct decision. After experimental validation, our system is capable of applying in the vehicle autonomous navigation and monitoring of agriculture. With the adaptive Bhattacharyya distance and cross matching methods, the experimental results indicate that the performance of target tracking is over 95%, and the trajectory prediction error is about 25 cm. Besides, local path planning technique can be conducted in real time vehicle navigation with 49% of map searching.

並列關鍵字

stereo vision ； Bhattacharyya distance ； Kalman filter ； A star

參考文獻

徐嘉鴻。2011。大尺度虛擬實境場景接合與修補演算法之研究。碩士論文。臺北：國立臺灣大學生物產業機電工程所。

賴宗誠。2012。應用多組雙眼攝影機系統進行車前三維環境模型重建。碩士論文。臺北：國立臺灣大學生物產業機電工程所。

Barth, A. and U. Frank. 2009. Estimating the driving state of oncoming vehicles from a moving platform using stereo vision. IEEE Transactions on Intelligent Transportation Systems. 10(4): 560-571.

Barth, A. and U. Frank. 2010. Tracking oncoming and turning vehicles at intersections. International IEEE Conference on Intelligent Transportation Systems (ITSC). 861-868.

Boykov, Y. Y. and M. P. Jolly. 2001. Interactive Graph Cuts for Optimal Boundary & Region Segmentation of Objects in N-D Images. Proceedings of Internation Conference on Computer Vision. 105-112.