Research On Key Technologies For Object Recognition Based On Depth Sensor

Three-dimensional object recognition technology is of great significance for the research of autonomous driving and intelligent robots.Its main mask is to recognize objects in the scene,and to locate the position of every object.In the research of intelligent robot,the object recognition is helpful to the robot to navigate and recognize objects in unknown environment.However,due to the diversity and complexity of the environment,three-dimensional object recognition is facing new challenges.Therefore,a new method of three-dimensional object recognition is proposed in this paper,which is based on a depth sensor.This contribution of this paper is as follows:A new three-dimensional descriptor—NSHOT is proposed.While the three-dimensional object is recognized,three-dimensional descriptors of points will be computed,and many descriptors required the local reference frame,such as PFH,FPFH and SHOT which are used widespread.However,with the fast incensement of the scale of points,a costly reference frame computation would need to be performed at every point of the cloud.Therefore,this paper presents a new method for 3D object recognition based on a descriptor without local reference frame computations.The experiment shows that this method is helpful to accelerate the process of three-dimensional object recognition and it will not affect the precision.A new method for point cloud surface matching based on NSHOT is proposed,the main work is as follows: Firstly,the NSHOT descriptors are extracted for all points of the model point cloud and the scene point cloud.Secondly,descriptors are compared using the Euclidean distance,and correspondences are set between each scene descriptor and its nearest neighbor(NN)in the model database by the Euclidean distance.Then,local model surfaces and local scene surfaces are defined for all the points from the correspondences.Lastly,a lot of transformations are computed by aligning the local model surfaces with the local scene surfaces corresponding to the correspondences.From the transformations,we should select the best transformation as the transformation between the whole model point cloud and the whole scene point cloud.The experiment shows that the rough matching is helpful to reduce the number of point correspondences to reduce the iteration,and it can speed up the process of the registration and 3D object recognition.