3D keypoint detectors and descriptors for 3D objects recognition with TOF camera

The goal of this work is to evaluate 3D keypoints detectors and descriptors, which could be used for quasi real time 3D object recognition. The work presented has three main objectives: extracting descriptors from real depth images, obtaining an accurate degree of invariance and robustness to scale and viewpoints, and maintaining the computation time as low as possible. Using a 3D time-of-flight (ToF) depth camera, we record a sequence for several objects at 3 different distances and from 5 viewpoints. 3D salient points are then extracted using 2 different curvatures-based detectors. For each point, two local surface descriptors are computed by combining the shape index histogram and the normalized histogram of angles between the normal of reference feature point and the normals of its neighbours. A comparison of the two detectors and descriptors was conducted on 4 different objects. Experimentations show that both detectors and descriptors are rather invariant to variations of scale and viewpoint. We also find that the new 3D keypoints detector proposed by us is more stable than a previously proposed Shape Index based detector.