3D Object Detection Based On Kinect

Along with the invention of RGB-D cameras such as kinect,the robot's vision ability has developed a lot.They can obtain not only the texture,but also the depth of the world.This brings new challenges to detect objects in 3D format.This paper mainly discussed the problem of 3D object detection.In location part,we using the spatial information of point cloud instead of sliding windows.In recognition step,this paper designed a new algorithm to extract point cloud feature instead of artificial feature.The main work of this paper is as follows.(1)In our daily life,the objects usually are placed on a horizontal support surface.So this problem can be solved in two steps.The detection of support plane and the cluster of objects.This paper first propose the distance cluster to segment the objects.In order to solve the overlapping situation,this paper proposes a new cluster method which based on supervoxels cluster.(2)In recognition step,this paper designed a new algorithm to extract point cloud feature.We first convert the point cloud of objects to depth map,then we run unsupervised learning algorithm such as K-Means to learn features from random patches.The learned features can be used as the CNN filters and convoluted over the input image to extract convolution feature.At last,we discuss two cluster methods in different situation.We also test our recognition method on two datasets.The results show that feature learned by single layer CNN can achieve higher recognition rate than artificially designed feature.