Research On Object Recognition And Localization Technology Based On 3D Point Cloud

The recognition of 3D objects in complicated scenarios is a quite challenging task in 3D computer vision.In the past few decades,since the acquisition of 2D images is very convenient,the recognition algorithm based on two-dimensional perception has been widely studied and has become a relatively mature field.With the rapid development of sensor technology,storage media and computer technology,more and more low-cost 3D sensors have emerged,making object recognition based on 3D data a research highlight in recent years.The 3D object recognition algorithms are mainly divided into two categories: global feature based and local feature based.Since the global features are very sensitive to occlusion and cluttered scenes,this paper systematically analyzes the 3D recognition algorithm based on local features.Then we propose an effective and robust Spatial Geometry Histograms(SGH)feature descriptor based on our improved local reference frame(LRF)method,aimed at robustly resisting the impacts of noise,varying mesh resolutions,occlusion and clutter.In addition,due to that the mainstream 3D feature detection algorithms are not capable to recognize texture-less objects in the industry,we propose a fast 3D recognition and pose estimation algorithm using planar patches for resolving the problem.In this paper,the object recognition and location technology based on 3D point clouds is deeply studied.The main research contents and the research results obtained include the following aspects:(1)For the repeatability and robustness of LRF,we propose a novel LRF algorithm.Different from other existing LRF proposals,we construct an auxiliary z'-axis for x-axis determination.The z-axis and z'-axis of LRF are calculated by different weighted covariance matrix on disparate radius neighbors of the keypoint respectively.X-axis is given via the cross-product of z-axis and z'-axis.The y-axis of the LRF is derived from the cross product of the z-axis and the x-axis.Furthermore,a scale factor is proposed for improving the robustness of LRF to mesh decimation,which can also be applied to other LRFs that suffer the same dilemma.Experimental results show that the algorithm achieves highly repeatability and robustness.(2)For the descriptiveness and robustness of local feature descriptors,a novel feature descriptor combining spatial distribution and geometric attributes,namely Spatial Geometry Histograms(SGH),is proposed.To achieve a prominent performance in terms of descriptiveness and robustness,the partitions along the radial and azimuth axes counting the spatial distribution of neighboring points,and the deviation angles between axes representing the geometric properties between adjacent points,are adopted to construct the descriptor.To evaluate the performances of our presented proposals,we conduct a series of experiments on eight public datasets compared with several state-of-arts methods.The results demonstrate our approaches achieve the integrally utmost performance indifferent application scenarios including shape retrieval,3D object recognition and 3D registration.(3)For the problem of difficult recognition and location of industrial texture-less objects,a 6DoF pose estimation algorithm based on geometric edges is proposed.Since the partial surface of most industrial parts is made up of planes,a simple and effective description method is constructed using planar geometric features.First,we perform an operation of detecting plane and then extract the corresponding edge image of plane by means of projection from 3D to 2D representations.Then we utilized the fast directional chamfer matching(FDCM)algorithm mentioned in [60] and improved its cost function,making it more robust to heavy occlusion.The final 6DoF pose estimation is resorted by the detection and localization of 2D instances.To evaluate the performances of presented proposals,we conduct a series of experiments on a number of synthetic and real sequences and compare it to state-of-the-art approaches based on 3D feature.The results demonstrate our approaches achieve the utmost performance in the presence of noise,clutter,and occlusions in terms of recognition rates and efficiency.