Title | ||
---|---|---|
Non-pre-trained Mine Pedestrian Detection Based on Automatic Generation of Anchor Box. |
Abstract | ||
---|---|---|
Mine pedestrian detection is an important part of computer vision and one of the key technologies of unmanned locomotive. In order to improve the structural adaptability of pedestrian detection network, reduce the workload of pre-training and reduce the risk of “negative migration” brought by migration learning, a non-pre-training underground pedestrian detection network based on anchor box is proposed. Firstly, the network model of mine pedestrian detection is introduced, including non-pretrained backbone network and branch structure detection network. Among them, the non-pre-trained backbone network mainly adds BatchNorm operation to make the gradient more stable and smooth. Anchor location prediction branch and anchor shape prediction branch in the detection network work together to improve the regression accuracy of anchor box. Secondly, the loss function of network training is described and the training parameters are adjusted by weighted loss. Finally, the experimental results based on the video of Taoyuan and Xinji mine in Anhui province are given. The experimental data show that the proposed algorithm can still maintain 96.3% AP at a real-time processing rate of 24 FPS. Compared with the RefineDet512, the AP increases by 2.4%. |
Year | DOI | Venue |
---|---|---|
2020 | 10.1007/978-3-030-59019-2_14 | WASA |
DocType | Citations | PageRank |
Conference | 0 | 0.34 |
References | Authors | |
0 | 5 |
Name | Order | Citations | PageRank |
---|---|---|---|
Xing Wei | 1 | 5 | 2.79 |
Changguang Wang | 2 | 0 | 0.34 |
Haitao Zhang | 3 | 0 | 0.34 |
Shaofan Liu | 4 | 1 | 1.71 |
Yang Lu | 5 | 183 | 50.38 |