3DHoNR: A 3D object recognition using an efficient and fast low-dimensional 3D descriptor for a Real-time Application

Recognition methods for 3D objects, which depend on local features, play a key role in ensuring robust recognition. Moreover, training-free approaches that require only a sample for recognition prove beneficial in real-time applications such as robotics, as they do not involve extensive learning processes. This paper proposes a training-free method for object recognition, introducing a resilient and efficient 3D descriptor generated from local features around keypoints. The approach ensures robustness by converting 3D keypoints into a new space and computing features in the original space. These features are derived from the Histogram of Point Distributions (HoPD) and the Histogram of Neighbors Relationship (HoNR). We conducted experiments on three prominent 3D datasets—Bologna, FEVOR and Kinect. Our proposed technique demonstrated a remarkable 100 % recognition rate on the Bologna dataset and surpassed state-of-the-art methods on FEVOR dataset. Additionally, the proposed technique is well-suited for real-time applications due to its low computation time requirements for object detection.