基於鳥瞰轉換下運用光流估測車輛環場景物之高度

Using Optical Flow under Bird's-eye View Transform to Estimate the Height of Objects around a Vehicle

雷丰銓(Fung-Chan Lei)

指導教授 ： 駱榮欽

國立臺北科技大學/電資學院/電腦與通訊研究所/碩士(2011年)

https://doi.org/10.6841/NTUT.2011.00260

摘要

本論文提出一種在鳥瞰影像下，使用單攝影機能即時估測動態物體高度之方法，一般而言，在一般影像平面下，無法直接從二維影像上獲得三維資訊(物體深度)，單攝影機需藉由額外資訊(如.攝影機角度與架設高度已事先知道)，雖然雙攝影機的深度圖被使用來估測物體深度，但它不方便運用在車上估測景物高度。因此，在單攝影機下運用光流估測物體高度被提出，在動態鳥瞰影像下，影像有兩種特色:第一，光流值與物體高度成正比，第二，在不同層高度的影像平面也無透視效應，由於攝影機至地面高度為已知的條件，再利用地面特徵點的光流與不同高度特徵點的光流比例值，並計算出特徵點的高度。為了驗證所提出的方法，我們建造了一個實驗平台，此實驗平台共有三層，每一層用不同顏色LED表示，紅色LED高度0公分(即為地面)，黃色LED高度30公分，綠色LED高度60公分，經由實驗平台往攝影機方向每平移10公分拍攝一次，獲得連續的鳥瞰影像，計算實驗平台上不同高度LED的光流，並估測LED高度，實驗結果顯示所計算的LED高度與實際LED架設高度非常接近，誤差率在±5%。

關鍵字

鳥瞰圖 ； 光流 ； 估測高度 ； 攝影機校正 ； 影像處理

並列摘要

The paper proposed a new method that can real-time estimates the height of objects with single camera from bird’s–eye view. Generally, it is impossible to obtain 3-D information, like the depth of objects, with one single camera except the additional information, such as the height of camera and the tilt angle of the camera, are known in advance. Although, the disparity map of the binocular cameras is usually employed to estimate the depth, but it is not suitable for a car to estimate the height (same as the depth estimating from planar view) of objects from bird’s-eye-view due to the difficulties of installation and corresponding. Therefore, the optical flow to estimate the height of the object with one camera is proposed. There are two features under dynamic bird’s–eye view of image. First, the optical flow value is proportional to the height of object. Second, there is no perspective effect in each layer height of image plane. Since the feature points between the obstacles and the ground are fetched and numbered, the ratio of optical flow of these 2 points is obtained. The camera’s height is known as well, therefore the height the of feature points can be estimated. In order to prove the proposed method, first, an experimental platform has been built, that includes one camera fixed on the platform and one 3*3 LEDs device. The LED device consists of three color LEDs where the height of red LED is 0 cm (i.e., on the ground), the height of yellow LED is 30 cm, the height of green LED is 60 cm. By using the camera, the continuous images of LEDs device per 10 cm movement under the bird’s-eye-view image is captured. From the captured images, the optical flows of LEDs (Red, yellow, green) in different height are calculated, then the heights of LEDs can be estimated. From the experimental results, the estimated LED’s height is very close to the actual LED’s height (5% error rate).

並列關鍵字

Optical Flow ； Estimate Height ； Camera Calibration ； Image Processing

參考文獻

[1] 曾定章,先進安全車輛多樣性視覺偵測辨識技術發展, 2011 Conference on Computer Vision Image Processing and Information Technology

[3] J. Salvi, J. Pagès, and J. Batlle, "Pattern codification strategies in structured light systems," Pattern Recognition, 2004, pp.827-849.

[4] D. Scharstein and R. Szeliski, "A taxonomy and evaluation of dense two-frame stereo correspondence algorithms," International Journal of Computer Vision, 47(1/2/3):7-42, April-June 2002.

[6] R.Y. Tsai, "A Versatile Camera Calibration Technique for High-Accuracy 3D Machine Vision Metrology Using Off-the-Shelf TV Cameras and Lenses," IEEE J. of Robotics and Automation, Vol. RA-3, No. 4, Aug. 1987, pp. 323-344.

[7] B. Wei, G. and De Ma, S., "Implicit and explicit camera calibration: theory and experiments," IEEE Trans. on PAMI, Vol. 16(5), 1994, pp. 469-479.