Real-time 3D object detection in unstructured environments

Abstract

Real-time specific 3D object detection plays an important role in intelligent service robot or intelligent surveillance fields. A novel discriminative learning based method is proposed to detect a specific 3D object in unstructured environments with a single RGB image. This method contains two stages: the candidate extraction stage and the recognition stage. On the candidate extraction stage, we use a simple, fast, and high quality objectness measure Binarized Normed Gradients (BING) to highlight the target candidate regions. On the recognition stage, each candidate region is verified by the designed cascade classifiers in order to be classified into different classes based on multi features including color, shape and texture contents. Our method is evaluated by its performance on our proposed challenging new dataset consisting of 3 objects and is compared in two public challenging dataset with other approaches for single RGB image based 3D object detection. The experiment results show that the proposed method can not only achieve a high precision (97%) at 90% recall, but also can meet the real-time processing requirement of about 22 fps on video sequences.