基於KINECT的室內環境偵測系統

Indoor Environment Detection System Based on Kinect Sensor

簡湘潾(Hsiang-Lin Chien)

指導教授 ： 余松年

國立中正大學/工學院/電機工程學系/碩士(2013年)

摘要

本研究的目的是使用Kinect 感應器，建構一個幫助盲人或移動機器人能夠
在室內環境中熟悉環境的室內環境偵測系統。
我們使用成本不高的Kinect 視覺感應器，以重建三維點雲來進行環境中的
平面偵測並以二維深度圖來進行樓梯偵測。在平面偵測方面，首先建構三維點雲，
然後進行點雲壓縮來減少點雲量並均勻點雲分布；接著帶入RANSAC 演算法並
且加以分割平面；然後再進行法向量估計，將所有分割的平面進行分類，以偵測
出環境中所有垂直於地面的大型平面障礙物。而樓梯偵測則是將所擷取到的深度
資訊中距離零到四公尺轉換為零到二五五的深度圖。主要步驟是先將感應器得到
的深度圖進行前處理以降低雜訊；接著為了偵測出圖中所有的直線，先進行邊緣
偵測，再用霍夫線偵測；最後將所有偵測出的直線進行一連串的判斷，找出所需
的直線並且換算成距離後輸出。
透過電腦的模擬，我們證實了本演算法能夠有效的偵測出地面與垂直地面的
大型平面障礙物；也能分辨上樓梯與下樓梯。計算感應器與平面的距離或是與上
下樓梯的直線距離都有一些誤差，然而都在可接受範圍，其中在感應器的雜訊較
多與偵測距離較遠時誤差會比較大。在處理速度上，在點雲中的平面偵測整體上
是比較耗時的，處理速度在個人電腦上一個場景能減少到一秒內，但在筆記型電
腦一個場景要花到十秒左右；而深度圖中的樓梯偵測由於是基於圖像處理，因此
運算量小並且在筆記型電腦就能夠作即時的處理。

關鍵字

平面偵測 ； 霍夫線偵測 ； 邊緣偵測 ； RANSAC ； 點雲 ； Kinect ； 樓梯偵測

並列摘要

The main objective of this study is to construct an environmental detection
system to help the blind or mobile robots to be familiar with the indoor environment.
In this study, we construct the 3D point cloud and 2D depth map to detect the
plane and staircases in the environments using the low-cost Kinect sensor. In plane
detection, we firstly downsample the data, reduce the amount of points and make the
cloud data more uniformly distributed after the construction of the 3D point cloud.
Then we process the data by using the RANSAC algorithm and estimate the normal
vector to classify all the detected plane. All of the planes from the environment that
are perpendicular to the ground will be detected. In order to detect the staircases, we
firstly transform the detected depth distance in the range of 0 to 4 meters into depth
map with values between 0 and 255. The noise are reduced by the proposed
pre-processing algorithm. Edge detection and Hough line detection follow to find all
the straight lines in the depth map. Finally, a series of judgment are applied to all the
detected straight lines to detect the lines we are interested and these distances are
calculated as outputs.
Through computer simulation, we have proved the capability of our algorithm in
efficiently detecting the ground planes and the huge planes that are perpendicular to
the ground. It can also distinguish the staircases of going upstairs or going downstairs.
Although some errors exists in the estimated distance between the sensor and the
plane or the distance between the sensor and the detected lines, they are within the
acceptable levels. However, the errors seem to increase with the level of sensor noise
and the distance detected. It is usually take longer time in processing point clouds.
The processing speed is no more than 1 second on the Desktop PC but increased to
around 10 second on the Laptop PC to process a frame. Comparatively, staircase
detection using depth map is solely based on image processing, thus less calculation is
needed and can be processed in real-time even on the Laptop PC.

並列關鍵字

plane detection ； Hough line detection ； edge detection ； RANSAC ； point cloud ； Kinect ； staircase detection

參考文獻

pp. 702-708, 2008.

[2] C. Zheng, R. Green, "Feature Recognition and Obstacle Detection for Drive

Zealand International Conference, 2011.

Mobile Navigational Aid for Visually Impaired Based on the Microsoft Kinect,"

“Blind Navigation with a Wearable Range Camera and Vibrotactile Helmet,”