A Lightweight RGB-D Image-Based 3D Object Detector for Mobile Robots

3D object detection enables mobile robots to operate effectively in real-world scenarios. To enhance the performance of 3D object detection without compromising inference speed, we present a novel 3D object detection method that leverages RGB-D images. Our proposed approach involves multiple stages. Firstly, we extract 2D bounding boxes from images and corresponding point clouds within frustums. Subsequently, the point clouds within the frustums are processed by a devised proposal generation module, generating bounding box proposals. Additionally, we leverage a refinement network to enhance the accuracy of 3D bounding box estimation. Our object detection framework requires minimal prior information and achieves a nearly real-time performance of 20 FPS while maintaining high accuracy. To validate the effectiveness of our proposed method, we conduct experiments on the SUN RGB-D 3D detection benchmarks. The results demonstrate that our approach outperforms other baseline methods regarding object detection accuracy, spatial complexity, computational cost, and inference speed.