摘要
针对室内RGB-D场景中3D目标检测对复杂背景的适应性较差、难以进行有效采样,以及场景推断时间较长等问题,本文提出一种基于融合采样策略的轻量级RGB-D场景3D目标检测方法。该方法以场景RGB-D数据作为输入,首先通过深度相机投影将其转化为三维点云场景;然后利用一种结合距离最远点采样和特征最远点采样的融合采样策略对场景点云进行采样,有效保留了场景各实例代表点,将所有特征代表点汇集在一起形成场景的特征代表点云;最后在代表点云中利用深度霍夫投票机制投票出场景各物体的中心,并对各物体周围的相关特征进行聚类,从而实现场景的3D目标检测。实验结果表明,与传统方法相比,所提框架的目标检测准确率得到有效提升,同一评估指标下的检测准确率平均提升2.1%,且同一环境下每个场景的推断速度仅需要57 ms,远快于传统方法2倍多,从而保证了室内场景3D目标检测的准确性和高效性。
Aiming at the problems of 3D object detection in indoor RGB-D scenes such as poor adaptability to complex backgrounds,difficulty in effective sampling,and long scene inference time,this paper proposes a lightweight 3D object detection method for RGB-D scenes based on fusion sampling strategy This method takes scene RGB-D data as input,and first converts it into a 3D point cloud scene through depth camera projection;Then,a fusion sampling strategy combining farthest distance point sampling and feature farthest point sampling is used to sample the scene point cloud,effectively preserving the representative points of each instance of the scene,which are collectively referred to as feature representative points;Secondly,all feature representative points are gathered together to form a feature representative point cloud of the scene;Finally,a deep Hough voting mechanism is used to vote on the center of each object in the scene in the representative point cloud,and relevant features around each object are clustered to achieve 3D object detection in the scene The experimental results of network training on SUN RGB-D dataset show that the proposed framework effectively improves the accuracy of target detection compared to traditional methods,with an average improvement of 2.1%under the same evaluation index.Moreover,the inference speed for each scene in the same environment is much faster than traditional methods,which only requires 57 ms,thereby ensuring the accuracy and efficiency of 3D target detection in indoor scenes.
作者
单丰
SHAN Feng(School of Computer Science and Technology,Zhejiang Sci-Tech University,Hangzhou 310018,China)
出处
《智能计算机与应用》
2024年第4期68-75,共8页
Intelligent Computer and Applications
