Abstract | ||
---|---|---|
Infrared images are not affected by visible light and can be used to continuously track pedestrian targets at night and in other complex environments. Pedestrian tracking is of great significance in intelligent monitoring, military investigation and other fields. However, due to the deformation and size changes caused by their changes in the process of night pedestrian tracking, the difficulty of the pedestrian tracking task is increased. Because of the above problems, this paper studies how to extract more feature information from infrared images and proposes an improved SiamRPN++ deformable convolution target deformation infrared tracking algorithm. An improved ResNet-50 deformable convolution network model is constructed to increase the convolution kernel offset. The extracted infrared target features have higher discrimination. At the same time, a weighted loss training method is proposed for classification and regression branches. The classification score is more consistent with the regression accuracy. The experimental results show that the improved SiamRPN++ algorithm improves the success rate and accuracy by 3.7% and 4.0%, respectively, compared with the mainstream SiamRPN++ algorithm. The tracking performance is effectively improved. |
Year | DOI | Venue |
---|---|---|
2022 | 10.1109/ICAC55051.2022.9911118 | 2022 27th International Conference on Automation and Computing (ICAC) |
Keywords | DocType | ISBN |
infrared image,pedestrian tracking,siamese network,deformable convolution,weighted loss | Conference | 978-1-6654-9808-1 |
Citations | PageRank | References |
0 | 0.34 | 5 |
Authors | ||
4 |
Name | Order | Citations | PageRank |
---|---|---|---|
Liuqing Yang | 1 | 2238 | 189.17 |
Jianwei Ma | 2 | 0 | 0.34 |
Zhaoyang Zhao | 3 | 0 | 0.34 |
Chao Ma | 4 | 85 | 27.49 |