An anchor-guided 3D target detection algorithm based on stereo RCNN

被引：0
|
作者：
Cao J. [1 ]
Tao C. [1 ,2 ]
机构：
[1] School of Electronic and Information Engineering, Suzhou University of Science and Technology, Suzhou
[2] Suzhou Automotive Research Institute, Tsinghua University, Suzhou
来源：
Yi Qi Yi Biao Xue Bao/Chinese Journal of Scientific Instrument | 2021年 / 42卷 / 12期
关键词：
3D target detection; Instance segmentation; Key point detection; Sparse anchor point; Stereo vision;
D O I：
10.19650/j.cnki.cjsi.J2107801
中图分类号：
学科分类号：
摘要：
The current binocular 3D detection algorithm has the problem of slow online calculation speed due to a large number of anchor points to be selected. To address this issue, an anchor-guided 3D target detection algorithm is proposed, which is based on the stereo RCNN. This method is named as the FGAS RCNN. In the first stage, a probability map is generated for the left and right input images to generate sparse anchor points and corresponding sparse anchor boxes. The left and right anchors are used as the whole entirety to generate a 2D preselection box. The second stage is based on the key-point generation network of the pyramid feature network. The key-point heatmaps are generated by the information of these sparse anchor points. A 3D bounding box can be generated by combining the stereo regressor with these key-point heatmaps. The original image will lose pixel-level information after convolution. The instance segmentation mask generated by Mask Branch can be used to solve this problem. The 3D bounding box center depth precision can be improved by the instance segmentation mask and the instance-level disparity estimation. Experimental results show that the proposed method can reduce the amount of calculation while maintaining a high recall rate without any depth and position prior information input. Specifically, the mean average precision is 44.07% on 3D target detection with a threshold of 0.7. Compared with the stereo RCNN, the proposed method improves the average precision by 4.5%. Meanwhile, the overall running time of our method is 0.09 s shorter than Stereo RCNN. © 2021, Science Press. All right reserved.
引用
收藏
页码：191 / 201
页数：10
相关论文
共 23 条
  • [1] REN S, HE K, GIRSHICK R, Et al., Faster R-CNN: Towards real-time object detection with region proposal networks, IEEE Transactions on Pattern Analysis and Machine Intelligence, 39, 6, pp. 1137-1149, (2016)
  • [2] HE K, GKIOXARI G, DOLLAR P, Et al., Mask r-cnn, Proceedings of the IEEE International Conference on Computer Vision, pp. 2961-2969, (2017)
  • [3] BRAZIL G, LIU X., M3d-rpn: Monocular 3d region proposal network for object detection, Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 9287-9296, (2019)
  • [4] LIU SH X, ZH Q H, MA D K, Et al., Development of elevator traction sheave wear detection system based on monocular vision [J], Journal of Electronic Measurement and Instrumentation, 34, 9, pp. 55-61, (2020)
  • [5] LI P, CHEN X, SHEN S., Stereo r-cnn based 3d object detection for autonomous driving, Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 7644-7652, (2019)
  • [6] DONG F X, CAI J, XIE Y M., Joint calibration method of stereo vision and 3D laser system, Chinese Journal of Scientific Instrument, 38, 10, pp. 2589-2596, (2017)
  • [7] QI C R, LIU W, WU C, Et al., Frustum pointnets for 3d object detection from rgb-d data, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 918-927, (2018)
  • [8] ZHENG SH W, LI W H, HU J Y., Vehicle detection in traffic environment based on laser point cloud and image information fusion [J], Chinese Journal of Scientific Instrument, 40, 12, pp. 143-151, (2019)
  • [9] YANG B, LUO W, URTASUN R., Pixor: Real-time 3d object detection from point clouds, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 7652-7660, (2018)
  • [10] YE Y F, WANG J ZH, Stair area recognition in complex environment based on point cloud, Journal of Electronic Measurement and Instrumentation, 34, 4, pp. 124-133, (2020)
123