文献类型：期刊文献

英文题名：Fusion-Enhanced Feature Deblurring for 3d Single Object Tracking

作者：Zhuang, Yi[1]; Zhao, Haitao[1]; Zhao, Kaijie[1]

机构：[1] School of information Science and Engineering, East China University of Science and Technology, 130 Meilong Rd, Shanghai, 200237, China

年份：2022

外文期刊名：SSRN

收录：EI(收录号：20220222797)

语种：英文

外文关键词：Image enhancement - Textures - Tracking (position)

摘要：The extremely sparse and incomplete shape description is inevitable in 3D single object tracking (SOT) on raw point clouds. RGB Images can offer complete texture and color information as the complementation of raw point clouds. One of the key problems of the feature fusion of point clouds and images is the ''feature blurring" caused by cross-view transformation. In this paper, we exploit both point cloud and RGB image features for the target and the search area. Substituting for the common correlation, the proposed feature deblurring fusion network (FDFN) enables an effective way in 3D SOT. Specifically, we first propose BoxBeams. Differing from vertical PointPillars, BoxBeams discretizes the raw point clouds from the horizontal view to ensure the consistency with the direction of RGB images, the box-related representation makes it more informative to depict an object. A dynamic spatial-attention fusion module is built to fuse the two diverse inputs to generate more representative feature maps. Leveraging progressive sampling, FDFN can adaptively learn where to locate the object and extract regions of interest (RoI) instead of depending on feature comparison. We further use a refinement module to make a more reliable feature embedding for determining the final bounding box (BBox). Experimental results on the KITTI benchmark show large improvements on state-of-the-art methods in four different tested categories. FDFN performs 13.4%higher in terms of success and 14.2% higher in terms of precision. ? 2022, The Authors. All rights reserved.