基于Kinect可通行性区域识别

发布时间：2019-01-03 09:50

【摘要】：有效而又可靠的可通行性区域识别对移动机器人的导航具有重要的意义。而目前关于这方面的研究成果主要集中于传感器可以直接探测到的有限的空间中,对于开阔的环境,这仍是一个挑战。鉴于传统单一传感器具有各自的局限性,如立体视觉的算法复杂与“短视”,激光雷达的昂贵,而组合传感器对传感器之间的数据同步提出了很高的要求。本文提出基于Kinect传感器的自监督可通行区域识别方法,对Kinect能有效探测到的近距离区域,根据一定规则识别地面和障碍物并标记,并将这两类识别标签投影到对应的RGB图像空间,然后在图像空间提取视觉和识别标签的组合特征,训练分类器,并使用该分类器对远距离图像空间进行分类识别,最终得到整个图像空间的可通行性。本文研究内容主要包括两部分:第一部分为近距离Kinect传感器障碍物识别,这一部分介绍了Kinect的软硬件及深度信息获取原理,在完成标定和RGB配准之后,采用二维图像结合空间3维坐标的方法,识别地面和障碍物。第二部分为基于障碍物识别的远距离可通行区域识别,这一部分将第一部分中近距离区域识别类别标签投影到对应的图像空间,之后使用滑动窗口将图像窗口分块,并基于图像块提取颜色、纹理、几何相结合的组合视觉特征,将提取的组合特征与对应图像块的类别标签相结合形成训练数据,去训练本文的分类器--Fuzzy ARTMAP,最后利用得到的分类器去分类远距离图像空间,得到远距离图像的识别结果。本文最后通过在室内室外两个场景进行实验,验证了Kinect近距离检测算法的有效性以及基于图像空间远距离可通行区域检测的有效性。
[Abstract]:Effective and reliable identification of traversable region is of great significance to the navigation of mobile robots. However, the current research results mainly focus on the limited space that sensors can detect directly, which is still a challenge for open environment. Since the traditional single sensor has its own limitations, such as the complexity of stereo vision algorithm and "shortsightedness", the high cost of lidar, and the combination of sensors put forward a very high requirement for data synchronization between sensors. In this paper, a self-supervised passable area recognition method based on Kinect sensor is proposed. The ground and obstacle can be identified and marked according to certain rules for the close-range area which can be detected effectively by Kinect. The two kinds of recognition tags are projected to the corresponding RGB image space, then the combined features of the visual and recognition tags are extracted in the image space, and the classifier is trained, and the classifier is used to classify and recognize the remote image space. Finally, the traversability of the whole image space is obtained. This paper mainly includes two parts: the first part is the close distance Kinect sensor obstacle recognition, this part introduces the Kinect software and hardware and the depth information acquisition principle, after completing the calibration and the RGB registration, Two-dimensional images combined with three-dimensional space coordinates are used to identify the ground and obstacles. The second part is based on obstacle recognition. In this part, the close area recognition category label is projected to the corresponding image space, and then the image window is divided into blocks using sliding window. Based on the combined visual features of color, texture and geometry extracted from image blocks, the combined features are combined with the category labels of corresponding image blocks to form training data to train the classifier of this paper, Fuzzy ARTMAP,. Finally, the remote image space is classified by using the classifier, and the recognition results are obtained. Finally, the effectiveness of the Kinect close-range detection algorithm and the detection of remote passable area based on image space are verified by two experiments in indoor and outdoor environments.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242

【参考文献】