Abstract
AbstractIn this paper, an improved 3D objective detection algorithm called Frustum PSPNet (F-PSPNet) is proposed combing RGB algorithm and point cloud data. First, the 2D region proposals in the RGB image are generated by using the 2D CNN object detector, and each obtained 2D region proposal is lifted to 3D frustum through the camera projection matrix. Next, the points in the frustums are grouped by sliding along the frustum axis with stride. To improve the utilization of feature information, each set of points is sent to the PointNet added attention mechanism to extract point-wise features. Then, F-PSPNet arrays these features as a feature map for use of its subsequent component of the Pyramid Scene Parsing Network (PSPNet). Finally, 3D object detection and class confidence are done in the proposed F-PSPNet. Experiments on KITTI dataset show that the detection accuracy of F-PSPNet has improved compared with other models, especially in pedestrian detection.Keywords2D object detection3D object detectionPoint cloudRGB imagesFrustumAttention mechanism
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have