A novel method for target detection and tracking
With weapons development currently emphasizing distance and automation, the shooting-range test has acquired heightened significance. Current issues include background contrast and target detection at various distances. In the past, films and videotaped images were used, and targets were labeled manually. But such methods are labor-intensive, detection is slow, and accuracy is low. To resolve these problems, we suggest a novel method, based on digital image processing, that can improve both efficiency and target precision. We employ real-time digital image capture, large-capacity storage, and automatic machine detection.
Our method may be briefly described as follows. First, gray pretreatment improves the contrast ratio in images of the target and background to be captured. Then, for effective segmentation, we employ the Otsu algorithm to create a binary image that separates target from background.1 Candidate targets are labeled in accordance with region-labeling rules. We compute the candidate targets and then compare the feature parameters with the standard target parameters to locate the genuine target. Finally, a projection algorithm is employed to detect its cross-center. Figure 1 charts the flow of target detection.
- Between-class variance: s B =ω0(µ0-µ Y )2+ω1(µ1-µ Y )2
- Total variance:
- Discriminant criterion:
- Where p(i) is a probability distribution=
- ω0 is a probability of class occurrence
- ω1 is a probability of class mean levels=1-ω0
- is the zeroth-order cumulative moment of the histogram up to the kth level
- is the first-order cumulative moment of the histogram up to the kth level
- is the total mean level of the original image.
- Optimal thresholds of an image depend on maximizing s B , so as to maximally increase the separability of the resulting classes of gray levels.
After target and background are roughly separated, the region-labeling method is employed to locate target position. The linked image pixel is labeled as a mark, or connected component, after which regional characteristics can be analyzed. The target position can be adjudicated by analyzing the characteristic features of the image, including area, perimeter, round degree, and similar values. Comparison of the candidate with the standard target in the image for maximal match enables the real target to be selected. Finally, its position is located, and the projection algorithm is used to detect the cross-center.2
Speed is important in real-time target detection. With precision a prerequisite, dynamic target tracking must be adopted on the basis of a static detection algorithm. A multistep image-detection algorithm has also been suggested.3 This would involve, first, constructing a pyramid of multilevel images.4 The original image would be divided into several same-size regions, each of a*a dimension. Each pixel located at the top left of the image region will be picked up. Consequently, these pixels will make up a new image through a process of delamination, carried out L times, developing images from high to low resolution. Suppose image S of N*N dimension, and the k layer image:
When k=0, the new image resolution is highest; when k=L, it is lowest. Figure 2 shows construction of a multilevel image pyramid.
Target detection starts from the lowest-resolution image to obtain a rough position from which higher-resolution efforts can be launched. Results indicate that the target position in the original image can be quickly and precisely detected. Experiments carried out on the shooting range indicate that the strategy works for dynamic target detection and tracking, and, as shown in Figure 3, that the algorithm achieves satisfactory results.
Lingjia Gu is a graduate student in the College of Electronic Science and Engineering at JiLin University. Her field of research is digital image processing and virtual instrumentation.
