| 引用本文: | 侯春萍,张倩文,王晓燕,王致芃.轮廓匹配的复杂背景中目标检测算法[J].哈尔滨工业大学学报,2020,52(5):121.DOI:10.11918/201907103 |
| HOU Chunping,ZHANG Qianwen,WANG Xiaoyan,WANG Zhipeng.Object detection in cluttered background based on contour matching[J].Journal of Harbin Institute of Technology,2020,52(5):121.DOI:10.11918/201907103 |
|
| 摘要: |
| 为解决基于形状的目标检测算法受图像复杂背景的影响,本文提出了一种新的基于轮廓匹配的复杂背景中目标检测方法,算法结合了显著性检测和模板匹配的方法.首先对输入图像在超像素级别进行预处理,应用显著性区域检测方法得到不含复杂背景的区域图像,然后在显著性区域内得到初始边缘图像,对初始边缘图像进行优化处理后利用形状描述子进行轮廓匹配,最后,通过深度优先的搜索策略识别目标的假设位置并进行假设验证来确定最终的目标位置,完成复杂背景图像中的目标检测任务.在ETHZ形状数据集的实验结果证明了本文算法的可行性,根据50%-IoU和20％-IoU标准与其它几种基于形状的目标检测方法进行对比,当误报率为0.3时,算法平均检测率是96%,误报率为0.4时,检测率已经达到99%,如果接受更高误报率时检测率可达到100%,均高于其余几种算法.算法的实验和对比分析结果表明本文方法可以提高检测精度,具有明显的性能优势,为复杂背景中的目标检测提供了新的解决方法. |
| 关键词: 目标检测 复杂背景 显著性检测 形状描述子 轮廓匹配 |
| DOI：10.11918/201907103 |
| 分类号:TN911.73 |
| 文献标识码:A |
| 基金项目:国家自然科学基金(3,2) |
|
| Object detection in cluttered background based on contour matching |
|
HOU Chunping1,ZHANG Qianwen1,WANG Xiaoyan2,WANG Zhipeng1
|
|
(1.School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China; 2.Academic Journal Publishing Center, Tianjin University, Tianjin 300072, China)
|
| Abstract: |
| To effectively reduce the effects of cluttered background on traditional shape-based object detection methods, a novel object detection algorithm based on contour matching was introduced, which combines the methods of saliency detection and template matching. First, the input image was preprocessed at the super pixel level to obtain the saliency region map without background by saliency feature detection. Then, the edge detection algorithm was applied to get the edge image in the saliency region, and the shape descriptor was used for contour matching after optimizing the edge image. Finally, a depth-first search strategy was applied to identify the hypothetical location of the object and perform hypothesis verification to determine the final location of the object. The experimental result in the ETHZ shape dataset proved the feasibility of this algorithm. Compared with other shape-based methods under the 50%-IoU and 20%-IoU evaluation criteria, according to the data results, the average detection rate of different categories was 96% when the false positive per image (FPPI) was 0.3. The detection rate was 99% when the FPPI was 0.4, and the detection rate would reach 100% if higher FPPI is accepted, which were all higher than the other algorithms. The experimental and comparative analysis results show that the proposed method could improve the detection accuracy and had obvious performance advantages, providing a new solution for object detection in cluttered background. |
| Key words: object detection cluttered background saliency detection shape descriptor contour matching |
