Enhanced Histogram Feature Descriptor for Automated Point Cloud Registration

In this paper, we propose an enhanced feature descriptor based on point feature histograms for fully automated point cloud registration task. To eliminate the overflow problem of reciprocal weights, the distance between the query point and its neighbouring point is utilized to support a exponential weight for each feature signature. Moreover, widely used bilinear vote in 2-dimensional computer vision are assembled to feature formation process, which will decrease the bin number of histograms and therefore accelerate the registration. To validate the method, we use synthetically generated geometric primitives to show the discriminative abilities of the feature descriptor. Further tests on real scans present a run time comparison of existing descriptors and our method. The results show that our method outperforms existing descriptors on real-time performance, meanwhile get slightly better discriminative power.