Traversable Region Recognition Based On Kinect

The effective and reliable traversable region recognition is of great significance to the navigation of mobile robots.The current research on this area is mainly focused on the sensor can be directly detected in the limited space,for the open environment,this is still a challenge.In view of the traditional single sensor has its own limitations,such as stereo vision algorithm complex and "short-sighted",the lidar is expensive,and the combination poses a high demand for data synchronization between the sensors.This paper based on Kinect The selfsupervision of the sensor can be used to identify the area,Kinect can effectively detect the close area,according to certain rules to identify the ground and obstacles and mark,and these two types of identification labels projected to the corresponding RGB image space,then the combination of visual and identification tags in the image space is extracted,Training the classifier,and the use of the classifier for remote image space classification and identification,and ultimately get the entire image space can be traversable.The main contents of this paper include two parts: the first part is the close of Kinect sensor obstacle identification,this part introduces Kinect's hardware and software and depth information acquisition principle,after the completion of calibration and RGB registration,the use of two-dimensional image binding space 3-dimensional coordinate method to identify the ground and obstacles.The second part is based on the obstacle identification of longrange traversable region identification,this part will project the the close range recognition category label of first part to the corresponding image space,and then use the sliding window to block the image window,and based on the image block to extract the combination of color,texture,geometry combined visual characteristics,Combining the extracted combination features with the category tags of the corresponding image blocks forms the training data,to train the classifier-Fuzzy ARTMAP,and finally use the resulting classifier to classify the long-range image space,The recognition result of the remote image is obtained.At last,through the experiments in indoor and outdoor scenes,the validity of Kinect close-range detection algorithm and the effectiveness of the image-based long-distance traversable region detection are verified.