融合FHOG和LBP特征的尺度自适应相关滤波跟踪算法 
下载： 995次
Scale-Adaptive Correlation Filter Tracking Algorithm Based on FHOG and LBP Features
华北理工大学电气工程学院, 河北 唐山 063210
图 & 表
图 1. FHOG特征提取示意图
Fig. 1. Schematic of FHOG feature extraction
下载图片 查看原文
图 2. LBP特征提取原理图
Fig. 2. Extraction principle of LBP feature
下载图片 查看原文
图 3. 当前帧的目标跟踪与尺度估计示意图
Fig. 3. Schematic of target tracking and scale estimation for the current frame
下载图片 查看原文
图 4. 精确率对比图
Fig. 4. Comparison of accuracy rate
下载图片 查看原文
图 5. 成功率对比图
Fig. 5. Comparison of success rate
下载图片 查看原文
图 6. 6种算法的部分跟踪效果图
Fig. 6. Diagrams of partial tracking effects of 6 algorithms
下载图片 查看原文
表 1选用的图像序列的信息描述
Table1. Information description of the selected image sequence
| Image sequence | Number of frames | Size /(pixel×pixel) | Property description |
|---|
| Bird1 | 408 | 720×400 | DEF, FM, OV | | Box | 1161 | 640×480 | IV, SV, OCC, MB, IPR, OPR, OV, BC, LR | | Skating2 | 473 | 640×352 | SV, OCC, DEF, FM, OPR | | Basketball | 725 | 576×432 | IV, OCC, DEF, OPR, BC | | Soccer | 392 | 640×360 | IV, SV, OCC, MB, FM, IPR, OPR, BC | | CarScale | 252 | 640×272 | SV, OCC, FM, IPR, OPR |
|
查看原文
表 2本文算法的主要参数设置
Table2. Main parameters of the algorithm in this paper
| Algorithm parameter | Value |
|---|
| Regularization parameter λ | 10-4 | | Ratio of searching areas | 1.6 | | Learning ratio of apparent model | [0.01,0.015] | | Number of directions of FHOG | 9 | | Cell size of FHOG /(pixel×pixel) | 4×4 | | Learning rate η | 0.01 | | Scale factor a | 1.02 | | Scale space N | 33 |
|
查看原文
表 36种算法在6组测试图像序列上的CLE
Table3. CLE of six algorithms on six sets of test image sequences
| Imagesequence | CLE /pixel |
|---|
| LCT | Staple | Struck | CSK | KCF | Our |
|---|
| Bird1 | 6.925 | 4.599 | 8.498 | 6.031 | 9.778 | 4.375 | | Box | 5.984 | 6.721 | 6.196 | 5.633 | 6.756 | 5.376 | | Skating2 | 6.969 | 3.095 | 3.065 | 2.606 | 5.189 | 2.482 | | Basketball | 9.765 | 7.386 | 7.352 | 7.269 | 12.889 | 7.198 | | Soccer | 4.033 | 4.93 | 4.356 | 4.017 | 3.997 | 3.949 | | CarScale | 6.927 | 6.511 | 6.281 | 5.975 | 10.684 | 5.918 |
|
查看原文
表 4不同算法的DP数据对比
Table4. Comparison of DP data of different algorithms
| Imagesequence | DP /% |
|---|
| LCT | Staple | Struck | CSK | KCF | Our |
|---|
| Bird1 | 100.00 | 34.90 | 36.50 | 35.20 | 95.00 | 100.00 | | Box | 92.80 | 86.40 | 87.20 | 91.60 | 86.40 | 100.00 | | Skating2 | 100.00 | 68.90 | 79.10 | 78.90 | 100.00 | 100.00 | | Basketball | 24.60 | 24.00 | 24.00 | 24.00 | 100.00 | 100.00 | | Soccer | 69.40 | 96.70 | 67.90 | 17.90 | 79.30 | 66.60 | | CarScale | 75.80 | 75.20 | 69.40 | 65.10 | 80.60 | 82.10 | | Mean | 77.10 | 64.35 | 60.68 | 51.17 | 90.22 | 91.45 |
|
查看原文
表 5不同算法的OP数据对比
Table5. Comparison of OP data of different algorithms
| Imagesequence | OP /% |
|---|
| LCT | Staple | Struck | CSK | KCF | Our |
|---|
| Bird1 | 100.00 | 27.60 | 26.80 | 27.60 | 36.40 | 100.00 | | Box | 30.60 | 46.40 | 52.30 | 63.20 | 74.20 | 62.80 | | Skating2 | 97.70 | 65.90 | 67.90 | 53.30 | 84.00 | 99.80 | | Basketball | 22.30 | 22.30 | 22.20 | 22.30 | 100.00 | 97.70 | | Soccer | 39.00 | 48.20 | 36.70 | 16.30 | 39.30 | 20.70 | | CarScale | 84.50 | 44.80 | 44.60 | 44.80 | 44.40 | 93.70 | | Mean | 62.35 | 42.53 | 41.75 | 37.92 | 63.05 | 79.17 |
|
查看原文
刘晓悦, 王云明, 马伟宁. 融合FHOG和LBP特征的尺度自适应相关滤波跟踪算法[J]. 激光与光电子学进展, 2020, 57(4): 041512. Xiaoyue Liu, Yunming Wang, Weining Ma. Scale-Adaptive Correlation Filter Tracking Algorithm Based on FHOG and LBP Features[J]. Laser & Optoelectronics Progress, 2020, 57(4): 041512.
PDF全文
