Bounding box initialization and optimization for ground target tracking in UAV Videos

Chuwei LI; Zhilong ZHANG; Ping ZHONG

doi:10.5768/JAO202344.0602003

Journal of Applied Optics, Volume. 44, Issue 6, 1332(2023)

Bounding box initialization and optimization for ground target tracking in UAV Videos

Chuwei LI... Zhilong ZHANG* and Ping ZHONG |Show fewer author(s)

National Key Laboratory of Science and Technology on Automatic Target Recognition, College of Electronic Science and Technology, National University of Defense Technology, Changsha 410073, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(17)

Fig. 1. Typical initial tracking box in physical systems

Download full size

View in Article

Fig. 2. Tracking results of different initialization methods

Download full size

View in Article

Fig. 3. Effect of square initial boxes with different deviations on tracking success value

Download full size

View in Article

Fig. 4. Effect of square initial boxes with different deviations on tracking precision value

Download full size

View in Article

Fig. 5. Strategy of tracking box initialization and optimization

Download full size

View in Article

Fig. 6. Global saliency map of UAV ground images

Download full size

View in Article

Fig. 7. Local saliency map of UAV ground images

Download full size

View in Article

Fig. 8. Comparison of image segmentation and salient region segmentation

Download full size

View in Article

Fig. 9. Flow chart of tracking box optimization algorithm

Download full size

View in Article

Fig. 10. Overall success plot and precision plot on VisDrone2018-SOT-test-dev dataset

Download full size

View in Article

Fig. 11. Overall success plot and precision plot on UAVDT dataset

Download full size

View in Article

Fig. 12. Limitations of tracking box optimization algorithms

Download full size

View in Article

Fig. 13. Role of distance normalization threshold

Download full size

View in Article

Fig. 14. Role of three-tier image pyramid structure

Download full size

View in Article

Table 1. Comparison of object tracking initialization methods

View table

View in Article

Table 1. Comparison of object tracking initialization methods

类别	方法	是否基于先验	速度	特点
自动	模板匹配^[5]	是	快	准确性低
自动	目标识别^{[6, 9]}	是	快	不具备通用性
自动	差分法^[7-8]	否	快	不适合复杂场景
自动	视觉显著性^[10-11]	否	快	不适合复杂场景
手动	任意长宽比矩形^[12]	否	慢	不适合运动场景
手动	正方形框	否	快	不适合狭长目标
半自动	正方形框+自适应优化（本文）	否	快	适合复杂场景，兼具通用性

Table 2. Algorithm pseudocode

View table

View in Article

Table 2. Algorithm pseudocode

算法1：基于视觉显著性和显著区域分割的跟踪框自适应优化算法
输入：原始图像 $I$ ，初始正方形跟踪框 $B$
输出：优化后的跟踪框 $B'$
1.根据初始正方形跟踪框 $B$ ，从原始图像 $I$ 中裁剪3个图像切片 $C_{1}$ 、 $C_{2}$ 、 $C_{3}$ （分别为正方形跟踪框的1.25、1.0和0.8倍大小）；
2.使用视觉显著性算法（BMS）对 $C_{1}$ 、 $C_{2}$ 、 $C_{3}$ 进行显著性检测，得到3幅显著图 $S_{1}$ 、 $S_{2}$ 、 $S_{3}$ ；
3.基于 $S_{1}$ 、 $S_{2}$ 、 $S_{3}$ ，分别使用二倍阈值分割和Saliency Cut对原始图像进行分割，得到6幅二值化掩模图像 ${M_{i}}_{i = 1}^{6}$ ；
4.将 ${M_{i}}_{i = 1}^{6}$ 映射到 $I$ ，计算对应的最小外接矩形位置 ${B_{i}}_{i = 1}^{6}$ ；
5.计算 ${B_{i}}_{i = 1}^{6}$ 的形心与 $B$ 中心的归一化距离 ${D_{i}}_{i = 1}^{6}$ ；
6.计算 ${B_{i}}_{i = 1}^{6}$ 的最长边与 $B$ 边长的比值 ${R_{i}}_{i = 1}^{6}$ ；
7. if $\exists {D_{i} T_{dist} & & R_{i} T_{ratio}}_{i = 1}^{6}$ then
8. $i^{*} = \underset{i}{\arg \max} (R_{i} \cdot (T_{dist} - D_{i}))$
9. $B' \leftarrow B_{i^{*}}$
10.else
11. $B' \leftarrow B$

Table 3. Parameters of bounding rectangle

View table

View in Article

Table 3. Parameters of bounding rectangle

外接矩形	在图像中的位置	形心与 $B$ 中心的归一化距离	最长边与 $B$ 边长的比值	综合得分
$B_{1}$	[622, 544, 680, 625]	0.0385	0.6923	0.1118
$B_{2}$	[615, 530, 680, 638]	0.0096	0.9231	0.1758
$B_{3}$	[616, 530, 679, 626]	0.0562	0.8205	0.1180
$B_{4}$	[622, 544, 680, 625]	0.0385	0.6923	0.1118
$B_{5}$	[632, 556, 668, 611]	0.0311	0.4701	0.0794
$B_{6}$	[617, 530, 678, 624]	0.0647	0.8034	0.1087

Tools

Get Citation

Copy Citation Text

Chuwei LI, Zhilong ZHANG, Ping ZHONG. Bounding box initialization and optimization for ground target tracking in UAV Videos[J]. Journal of Applied Optics, 2023, 44(6): 1332

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites