1. A Method of Erect Rail Barricade Recognition Based on Forward-Looking 3D Sonar
- Author
-
Bao-qi LI, Lu-lu REN, Fa CHEN, Bin QIAN, and Hai-ning HUANG
- Subjects
forward-looking three-dimensional sonar ,beamforming ,sonar image ,target detection ,attention mechanism ,Naval architecture. Shipbuilding. Marine engineering ,VM1-989 - Abstract
To alleviate the difficulty in detecting and recognizing erect rail barricades, a forward-looking three-dimensional(3D) imaging sonar is used to improve the effectiveness of erect rail barricade recognition, and a single shot detector(SSD)-based 3D point cloud(PCSSD) target recognition method is proposed. First, the original beam domain data are processed using a threshold filter and pass-through filter. Next, forward projection of the filtered 3D point cloud data is performed, and the grayscale depth image and pseudo color depth image are obtained. The SSD was then used to recognize the targets from the pseudo color depth image. Subsequently, the depth range of the selected targets is calculated based on the characteristics of the targets recognized from the grayscale depth image. Finally, erect rail barricades in the form of 3D point clouds were detected by combining the two-dimensional detection results and the corresponding depth range. A feature extraction block based on the multi-scale attention mechanism was proposed, and a novel target detector SSD-MV3ME was used for the design. On the 3D point cloud erect rail barricades dataset GTZ, under the condition of the same detection time, the detection accuracy of SSD-MV3ME was higher than that of the lightweight target detection model SSD-MV3 by 1.05%, and the model parameters had less than 2 482 kB. The results show that the PCSSD based on SSD-MV3ME is more suitable for erect rail barricade recognition tasks.
- Published
- 2022
- Full Text
- View/download PDF